UNIVERSITÉ DE LIÈGE Faculté des Sciences Appliquées Dynamic State Estimation of Synchronous Generators using Phasor Measurement Units. par Alberto DEL ÁNGEL HERNÁNDEZ Master of Science in Electrical Engineering, Instituto Politécnico Nacional, México Thèse présentée en vue de l’obtention du grade de Docteur en Sciences Appliquées. Année Académique 2006-2007
132
Embed
UNIVERSITÉ DE LIÈGE - uliege.beI am deeply grateful with Professor Mania Pavella, for her kind support received when I was ... I would like to express my gratitude to Philip Mack
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
UNIVERSITÉ DE LIÈGE
Faculté des Sciences Appliquées
Dynamic State Estimation of Synchronous Generators using Phasor Measurement Units.
par
Alberto DEL ÁNGEL HERNÁNDEZ
Master of Science in Electrical Engineering, Instituto Politécnico Nacional, México
Thèse présentée en vue de l’obtention
du grade de Docteur en Sciences Appliquées.
Année Académique 2006-2007
2
3
Acknowledgments First of all, I would like to thank my supervisor, Professor Louis Wehenkel, for his time,
patience, dedication, support and guidance during all my PhD studies. Also I thank for the
confidence that he had put in me during these last years of studies
Thanks to accept to be my promoter of thesis and thanks a lot for his experience, and
everything what I have learned in the field of research.
Thank you very much Dr Wehenkel.
I am deeply grateful with Professor Mania Pavella, for her kind support received when I was
arriving for the first time to Liege.
I also thank to Dr. Daniel Olguin-Salinas who gave me the opportunity to come to the
University of Liege to follow my PhD studies. Thanks for his help and advice during all the
time that I have stayed in Liege.
Thank to Dr. Mevludin Glavic for his help and support during my PhD studies.
I would like to express my sincere gratitude to Dr. Pierre Geurts, Dr. Damien Ernst and
Dr. Raphael Maree for their discussions, help and research collaborations.
I would like to express my gratitude to Philip Mack for the permission to use PEPITo
software, and all the staff of PEPITe for provides me their help kindly.
I thank to the Comision de Operación y Fomento de Actividades Academicas (COFAA), the
Programa SUPERA of ANUIES (Mexico) and the Instituto Politecnico Nacional of Mexico
for the financial support they granted me during the first years of my PhD studies.
I gratefully acknowledge the Comité National d’Accueil, Asbl. (Belgium) for the financial
support that I had receive during this last year.
I would like to thank to my parents for their support and also to my family for their patience
and understanding during this time far our homeland.
4
Summary
This main objective of this thesis is to develop and improve techniques for estimating
and predicting rotor angles, speeds and accelerations in the time frame of transient
(angle) stability of electric power systems. The investigated dynamic state estimation
technique is based on the use of voltage and current phasors that can be acquired in
real-time using a PMU (phasor measurement unit) located at the EHV bus of the
power system. The research is based on simulation data because techniques for the
direct measurement of rotor angle were not available.
The determination of security levels, for given operating conditions, traditionally has
been done using deterministic criteria. Under deterministic criteria, an operating
condition is identified as secure if it can withstand the effects of each and every
contingency in a pre-specified contingency set. Withstanding the effects means that the
given contingencies will not violate branch loading or nodal voltage criteria in steady
state conditions or make the system dynamically unstable [IEEE04].
The task of assessing the level of security for a given operating condition or topology
configuration, often leads to the definition of a security margin using some selected
variables or parameters. The choice of these variables or parameters depends on the type
of phenomena limiting the system. Given the high complexity of power systems and the
large range of time constants of its dynamics, the study of power system security has led
CHAPTER 2 BACKGROUND
23
to the decomposition of the notion of security into various sub-problems along different
criteria. Below, we introduce the main notions and definitions that have been introduced
in this field.
Power system stability. Stability is the ability of an electrical power system, for a given
initial operating condition, to regain a state of operating equilibrium after being subjected
to a physical disturbance, with most system variables bounded so that practically the
entire system remains intact [IEEE/CIGRE04].
The terminology about power system stability proposed in [IEEE/CIGRE04] is based on
the following considerations:
• The physical nature of the resulting mode of instability as indicated by
the main system variable in which instability can be observed.
• The size of the disturbance considered which influences the method of
calculation and prediction of stability.
• The devices, process and the time span that must be taken into
consideration in order to assess stability.
There are several main divisions in the study of power system dynamics and stability
[SP98]. De Mello classified dynamic processes into three categories:
• Electrical machine and system dynamics.
• System governors and generating control.
• Prime-mover energy supply dynamics and control.
Concordia and Schultz classify dynamics studies according to four concepts [SP98,
CS75]:
• The time of the system condition: past, present and future.
• The time range of the study: microsecond through hourly response.
• The nature of the system under study: new station, new line, etc.
24
• The technical scope of the study: fault analysis, load shedding, sub
synchronous resonance, etc.
All these classifications share a common thread: they emphasize that the system is not in
steady state and that many models for various components must be used in varying
degrees of detail to allow efficient and practical analysis [SP98].
Disturbance. A disturbance in a power system is a sudden change or sequence of
changes in one or more of the parameters of the system, or in one or more of the physical
quantities [PM94].
Small disturbance. A small disturbance is a disturbance for which the set of equations
that describe the power system may be linearized for purpose of analysis [PM94].
Large disturbance. A large disturbance is a disturbance for which the equations that
describe the power system cannot be linearized for the purpose of analysis [PM94].
Pre-fault system. A power system immediately preceding the initiation of a large
disturbance is termed a “pre-disturbance (pre-fault) system”. The system is usually
considered to be in steady state in this phase [PM94].
Fault-on system. In the during disturbance (or during fault or fault-on) system the power
system is under the continuous influence of a disturbance (or a sequence of disturbances);
this phase lasts for the entire duration of the disturbance. This is the initial stage of the
transient period [PM94].
Post-fault system: A power system immediately following the complete isolation of a
large disturbance is termed a “post-disturbance (post-fault) system”. During the post-fault
phase the transient period continues and the system may or may not eventually reach a
steady-state. The post-disturbance phase decides whether the system is stable or not
[PM94].
CHAPTER 2 BACKGROUND
25
Figure 2.2 gives the overall picture of the power system stability problem, identifying its
categories and subcategories.
Figure 2.2 Classification of Transient Stability. Adapted from [Kun94].
Classification of power security assessment and control can be defined as follows
[RV03]:
Static security assessment (SSA), are methodologies that verify bus voltage and line
power flow limits for the post-contingency steady state operating condition,
considering that the transition between the pre-contingency and the post-
contingency steady state operating states has taken place without suffering any
instability phenomena in any part of the system. Static security assessment
essentially verifies the existence of a post-fault steady state that satisfies all
constraints deemed important for this state to survive for a long enough period of
time.
26
Dynamic Security Assessment (DSA), are methodologies for evaluating the stability
and quality of the transient processes between the pre-contingency and post-
contingency steady states. DSA aims at ensuring that the system will be stable
after the contingency occurrence and that the transient caused by such
contingency will be well damped, of small amplitude and with little impact on the
quality of service. Dynamic security assessment essentially verifies that the
system will reach and remain in a neighborhood of its post-fault steady state
conditions.
The following are description of the corresponding forms of instability phenomena:
Voltage instability is the inability of a power system to maintain steady voltages at all
buses in the system after being subjected to a disturbance from a given initial
operating condition. It depends on the ability to maintain/restore equilibrium,
between load demand and load supply from the power system. Voltage
instabilities generally occur in the form of a progressive fall or rise of voltages
of some buses [Kun04].
Frequency instability refers to the inability of a power system to maintain steady
frequency following a severe system upset resulting in a significant imbalance
between generation and load. It depends on the ability to maintain/restore
equilibrium between system generation and load, with minimum unintentional
loss of load. Frequency instability generally occurs in the form of rapid
frequency drops leading to tripping of generating units and/or loads [Kun04].
Rotor angle instability is the inability of synchronous machines of an interconnected
power system to remain in synchronism after being subjected to a disturbance.
It depends on the ability to maintain/restore equilibrium between
electromagnetic torque and mechanical torque of each synchronous machine
in the system. Transient instability may occurs in the form of increasing
angular swings of some generators leading to their loss of synchronism with
CHAPTER 2 BACKGROUND
27
other generators [Kun04]. One often distinguishes between plant mode
instability, where a single power plant looses his stability and inter-area mode
instability where all the plants of large area loose their synchronism with
respect to the rest of the interconnection.
For convenience in analysis and for gaining useful insight into the nature of stability
problems, it is useful to characterize rotor angle instability in terms of the following two
subcategories [Kun04]:
Small-disturbance (or small-signal) rotor angle instability is concerned with the
inability of the power system to maintain synchronism under small
disturbances. In today’s power systems, small-disturbance rotor angle stability
problem is usually associated with insufficient damping of oscillations. It is
generally characterized by negatively damped power swings among remote
generators of the interconnection, typically leading to the tripping of
interconnection lines and/or generators.
Large-disturbance rotor angle instability or Transient Stability is concerned with the
inability of the power system to maintain synchronism when subjected to a
severe disturbance, such as a short circuit on a transmission line. The resulting
system response involves large excursions of generator rotor angles and
speeds, followed by generator tripping due to over or under speed protections.
And finally, perhaps the most import classification of dynamic phenomena is their natural
time range of response. A typical classification is shown in figure 2.2. This time-range
classification is important because of its impact on component modeling and on the
response speed of control and protective devices needed to counter the corresponding
instabilities. It should be intuitive that is not necessary to solve the complex transmission
line wave equations to investigate the impact of a change in boiler control set points. This
brings to mind the statement that “the system is not in steady state”. Depending on the
28
nature of the dynamic disturbance, portions of the power system can be considered in
“quasi-steady state” [SP98].
Figure 2.3 Time Ranges of dynamic phenomena. Adapted from [SP98]
2.1.4 The swing equations
This thesis is devoted to the estimation of rotor angles and speeds in the framework of
transient stability. We will therefore consider the behavior of the system immediately
following a disturbance such as a short circuit on a transmission line, the opening of a
line or the switching on a major load to name just a few.
Since a synchronous machine is a rotating body, the laws of mechanics applying to
rotating bodies apply to it [Kim64]. The equations of central importance in power system
transient stability analysis are the rotational inertia equations describing the effect of
unbalance between the electromagnetic torque and the mechanical torque of the
individual machines [Kun94]. A brief description of the establishment of the swing
equation is expressed below, along the lines given in [Kun94].
CHAPTER 2 BACKGROUND
29
When there is an unbalance between the torques acting on the rotor, the net torque
causing acceleration or deceleration is
ema TTT −= (2.3)
Where aT denotes the accelerating torque, mT is the mechanical torque and eT
electromagnetic torque. All units are N.m. and in the above equations, Tm and eT are
positive for a generator and negative for a motor.
The combined inertia of the generator and turbine is accelerated by the unbalance in the
applied torques. Hence, the equation of motion is:
emm TTTa
dt
dJ −==
ω , (2.4)
where J is the combined moment of inertia of generator and turbine; kg.m2, mω is the
angular velocity of the rotor, mech. rad/sec and t time in seconds.
Defining the inertia constant H as the kinetic energy in watts-seconds at rated speed
divided by the VAbase and denoting by m0ω the rated angular velocity in mechanical
radians per second, the inertia constant is obtained by
H = 12
Jω 20m
VAbase
. (2.5)
The moment of inertia J in terms of H is obtained by
basem
VAHJ
J0
2
2
ω= . (2.6)
Substituting the above equation in eq. (2.4) gives
emm
basem
TTdt
dVA
H −=ω
ω 02
2 , (2.7)
30
and rearranging yields
mbase
em
m
m
VA
TT
dt
dH
00 /2
ωωω −
=
. (2.8)
The equation of motion in per unit form is thus
emr TT
dt
dH −=
ω2 , (2.9)
where
ω r = ωm
ω0m
=ωr /Pf
ω0 /Pf
= ω r
ω0
, (2.10)
where rω is the electrical angular velocity of the rotor in electrical rad/sec, ω0 is its rated
value, and Pf is number of field poles.
If δ is the angular position of the rotor in electrical radians with respect to a
synchronously rotating reference and 0δ is its value at 0=t ,
δ(t) = δ0 + ωr (τ)d0
t
∫ τ −ω0t . (2.11)
In other words we have
rrdt
d ωωωδ ∆=−= 0 (2.12)
and
d2δdt 2
= dωr
dt=
d ∆ωr( )dt
= ω0
dω rdt
= ω0
d ∆ω r( )dt
.
(2.13)
CHAPTER 2 BACKGROUND
31
Substituting for dt
d rω given by the above equation in equation (2.9), we get
em TTdt
dH −=2
2
0
2 δω
. (2.14)
It is often desirable to explicitly isolate in these equations a component of damping
torque which has its origin in a linear dependence of the mechanical and/or
electromagnetic torques on the speed deviation. This leads to a slightly different swing
equation formulated as follows:
2H
ω0
d2δdt 2
= T 'm − T 'e − KD∆ω r (2.15)
Equation (2.15) represents the equation of motion of a synchronous machine. It is
commonly referred to as the swing equation because it represents swings in rotor angle δ
during disturbances.
The state-space form requires the component model to be expressed as a set of first order
differential equations. The swing equation (2.15), expressed as two first order differential
equations, becomes
( )rDemr KTT
Hdt
d ωω∆−−=
∆2
1 (2.16)
rdt
d ωωδ ∆= 0 (2.17)
In the above equations, time t is in seconds, rotor angle δ is in electrical radians, and 0ω
is equal to fπ2 . The block diagram form representation of equations (2.16) and (2.17) is
shown in figure 2.4.
32
DKHs +2
1
s
0ω
eT
rω∆mT
Figure 2.4 Block diagram representation of swing equation. Adapted from [Kun 94].
Characterization of the power system operating states. In his pioneering work, DyLiacco introduced the idea that the power system may operate
in the following modes: Normal, Alert, Emergency and Restorative [EPRI81,Dyl68].
More recently, Fink and Carlsen expanded this concept by identifying the constraints
satisfied or violated in each mode of operation. Three sets of generic equations (one
differential and two algebraic ones) govern power system operation: the differential set
encodes the physical laws governing the dynamic behavior of the systems components.
The two algebraic sets comprise ‘equality constraints’, which refer to the system’s total
load and total generation, and ‘inequality constraints’, which state that some system
variables, such as currents and voltages, must not exceed maximum levels representing
the limitations of physical equipment [FC78]. :
• Normal: all equality and inequality constraints are satisfied; reserve
margins are adequate to withstand stresses.
• Alert: all constraints are still satisfied; reserve margins are such that
some disturbance could result in a violation of some inequality
constraints.
• Emergency: some inequality constraints are violated; the system is still
intact and control actions could be initiated to restore the system to at
least the alert state.
• In extremis: equality constraints and inequality constraints are violated;
the system will no longer be intact and a portion of the load will be lost.
CHAPTER 2 BACKGROUND
33
• Restorative: control actions are being taken to pick up the lost load and
to reconnect the system.
Figure 2.5 shows the five different operating states of the power system.
SecureNormal
E: Equality Constraints
I: Inequality Constraints
- : Negation
Load Tracking, Cost Minimization,
System Coordination
Restorative Alert
In Extremis Emergency
E,I
E,I
E,I
E,I
- -
-
- E,I
Insecure
System Not Intact System Intact
Figure 2.5 Power system operating states. Adapted from [SC02]
34
2.2 STATE ESTIMATION AND AUTOMATIC LEARNING
2.2.1 Automatic learning Automatic learning (AL) is a term used to denote a highly multidisciplinary research field
of methods, which aim to extract information (knowledge) from databases containing
large amounts of low-level data. AL encompasses three main families of methods:
• Statistical data analysis and modeling,
• Artificial neural networks (ANN)
• Symbolic machine learning in artificial intelligence.
During the last 20 years, many researches worked in the topic, applying different
techniques (statistical pattern recognition, neural networks and machine learning) to
different power system problems (load forecasting, system identification and state
estimation, stability assessment and control) [Weh98].
Automatic learning methods essentially aim at extracting a model of a system from the
sole observation (or the simulation) of this system in some situations. By model, we mean
some relationships between the variables used to describe the system in some
encountered situations or to help understating its behaviour [Weh98].
2.2.2 Supervised learning Supervised learning is the part of automatic learning that focuses on modelling
input/output relationships. More precisely, the goal of supervised learning is to identify a
mapping from some input variables to some output variables on the sole basis of a sample
of observations of the values of these variables. The variables are often called (input or
output) attributes, observations are called objects and the sample of objects is the
learning sample. In the context of security assessment, an object would thus correspond
to an operating state of a power system, or more generally to a simulated security
scenario. The input attributes would be relevant parameters describing its electrical state
and topology and the output could be information concerning its security, in the form of
either a discrete classification (e.g. secure/insecure) or a numerical security margin
[Weh98].
CHAPTER 2 BACKGROUND
35
The general problem of supervised learning is formally stated as follows [Geu02]:
For any value of N and a learning sample LsN and without any a priori
knowledge of the functions P(.), Y(.), or, A(.), find a function f(.) defined on A
which minimizes the expected prediction error defined by:
Err f( )= EA ,Y L Y, f A( )( ){ }= L(y(o), f (a(o)))dP(o)U
∫ (2.18)
where U denotes the universe of all possible objects, y(.) the output attribute (a
function defined on U) and a(.) the vector of input attributes (another function
defined on U). L(.,.) is a loss function which measures the discrepancy between its
two arguments, P(.) is a sampling probability distribution defined over U and LsN
is a sample of N observed objects for which both y and a are given as inputs to the
supervised learning algorithm.
There are two main types of supervised learning problems:
• Classification problem: the output attribute takes a finite number of discrete
values.
• Regression problem: the output attribute takes a possible infinite number of real
values.
In this thesis, we are focused integrally to the regression problem.
2.2.3 Main classes of supervised learning algorithms
In this section, we provide a brief overview of the main types of supervised algorithms
that exist in the literature. In this dissertation we will mainly use ANNs and more
specifically MLPs. MLPs will be explained with more details in a subsequent section.
36
2.2.3.1 Linear regression
Linear regression is one of the oldest forms of machine learning. It is a long established
statistical technique that involves simply fitting a line (or a hyperplane) to some data. The
easiest case for linear regression is when the examples have a single numeric input
attribute and a numeric output value, i;e; if there are N examples, where the attributes for
each example is called ix and the label for each is iy . We can envision each example as
being a point in 2-dimensional space, with an x-coordinate of ix and a y-coordinate of iy .
Linear regression would seek the line ( ) bmxxf += that minimizes the sum-of-squares-
error for the training samples:
y i − f x i( )( )2
i=1
N
∑ . (2.19)
The quantity y i − f x i( ) is the distance from the value predicted by the hypothesis line to
the actual value – the error of the hypothesis for the training sample i. Squaring this value
gives grater emphasis to large errors and saves us dealing with complicated absolute
values in the mathematics while minimizing eq. (2.19) rtω m and b.
2.2.3.2 Decision trees Decision tree learning is a method for classification problems, in which the learned
function is represented by a decision tree. Learned trees can also be represented as sets of
if-then rules to improve human readability. Decision trees classify instances by sorting
them down the tree from the root to some leaf node, which provides the classification of
the instance. each node in the tree specifies test of some attribute of the instance, and
each brand descending from that node corresponds to one of the possible values for this
attribute. An instance is classified by starting at the root node of the tree, testing the
attribute specified by this node, then moving down the tree branch corresponding to the
value of the attribute in the given example. This process is then repeat for the subtree
rooted at the new node. [Mit97].
Figure 2.6 illustrates a typical learned decision tree.
CHAPTER 2 BACKGROUND
37
Figure 2.6 A decision tree for the concept Play Tennis. Adapted from [Mit97]
The decision tree shown in Fig. 2.6 corresponds to the expression:
( )( )
( )WeakWindRainOutlook
OvercastOutlook
NormalHumiditySunnyOutlook
=∧=∨=∨
=∧= (2.20)
A decision tree (DT) is obtained from a partitioning tree by attaching classes to its
terminal nodes. The tree is seen as a function; associating to any object the class attached
to the terminal node, which contains the object [Weh98].
The main strength of decision trees is their interpretability. By merely looking at the test
nodes of a tree one can easily sort out the most salient attribute and find out how they
influence the output. Another very important asset is the ability of the method to identify
the most relevant attributes for each problem. The last characteristic od DT is
computational efficiency: tree growing computational complexity is practically linear in
the number of candidate attributes and in number or learning states, allowing one to
tackle easily problems with a few hundred candidate attributes and a few thousand
learning states [Weh98].
38
Although variety of decision tree learning methods have been developed with somewhat
differing capabilities and requirements, decision tree learning is generally best suited to
problems with the following characteristics [Mit97] :
• Instances are represented by attribute-value pairs: Instances are described by a
fixed set of attributes and their values.
• The target function has discrete output values: the decision tree in Fig. 2.6 assigns
a Boolean classification to each example (i.e. yes or no).
• Disjunctive description may be required: decision trees naturally represent
disjunctive expressions.
• The training data may contain errors: Decision tree learning methods are robust to
errors, both errors in classification of the training examples and errors in the
attribute values that describe these examples.
• The training data may contain missing attribute values. Decision tree methods can
be used even when some training examples have unknown values.
2.2.3.3 Regression trees Regression trees may be considered as a variant of decision trees, designed to
approximate real-valued functions instead of being used for classifications tasks.
The inner nodes of regression trees are marked with test as in decision trees. The
difference is, that the leaves of regression trees may be marked with arbitrary real values,
whereas in decision trees the leafs may only be market with elements of a finite set of
discrete values. A further extension is to allow linear functions as label of leaf nodes. In
this case the function at the leaf node reached for a specific example is evaluated for the
instance’s attributes values, to determine the value of the target attribute. This allows for
global approximating by using multiple local approximations.
Regression tree induction is a well-known approach for improving along a continuous,
output dimension [BFOS84].
Regression trees decompose the attribute space into a hierarchy of regions. Similary to
decision trees, regression trees are built in a top-down approach: starting with the top-
node and the complete learning set, an attribute ia and a threshold value iv are selected
CHAPTER 2 BACKGROUND
39
to decompose the learning set into two subsets, corresponding to states for which i ia v<
and i ia v≥ respectively.
The procedure continues splitting until either the variance has been sufficiently reduced
or it is not possible to reduce it further in a statistically significant way [Weh96].
Figure 2.7shows a representation of a regression tree and an approximation of the
numerical output.
y
1a2α 1α
6
4
5.1
1a
1a
1a< 1a≥
2a<2a≥
6
45.1
Figure 2.7 Example of a regression tree. Adapted from [Ola04]
Regression trees have been used in fields as diverse as air pollution, criminal justice, and
the molecular structure of toxic substances. Its accuracy has been generally competitive
with linear regression. It can be much more accurate on non linear problems but tends to
be somewhat less accurate on problems with good linear structure [BFOS84].
2.2.3.4 Ensemble methods. Ensemble methods consist in growing several models with a classical machine learning
algorithm. Then, the predictions of these models are aggregated to provide a final
prediction potentially better than individual ones. One of the most popular family of
ensemble methods is defined by Perturb and Combine methods, that consist in perturbing
the learning algorithm and/or the learning sample so as to produce different models from
the same learning sample. The predictions of these models are then aggregate bya simple
average or a majority vote in the case of classification [Geu03].
40
Bagging
The Bagging Algorithm (Bootstrap aggregating) by Breiman [Bre96] votes classifiers
generated by different bootstrap samples (replicates). A bootstrap sample is generated by
uniformly sampling m instances from the training set with replacement. T bootstrap
samples 1 2, ,..., TB B B are generated and a classifier iC is built from each bootstrap
sample iB . A final classifier *C is built from 1 2, ,..., TC C C whose output is the class
predicted most often by its sub-classifiers, with ties broken arbitrarily.
Table 2.1 shown the Bagging algorithm how works:
Table 2.1 The Bagging algorithm. Adapted from [BK99]
{
Input: training set , Inducer , integer (number of bootstrap samples)
1. for 1
2. = bootstrap sample from ( i.i.d. sample with replacement)
3.
S T
i to T
S S
=′
I
}*
: ( )
*
( )
4.
5. ( ) arg max 1 (the most often predicted label y)
Output : classifier
i
i
y Y i C x y
C S
C x
C
∈ =
′=
= ∑
I
For a given bootstrap sample, an instance in the training set has probability ( )1- 1-1/m
m
of being selected at least once in the m time instances are randomly selected from the
training set. For large m , this is about 1 1/ 63.2%e− = , which means that each bootstrap
sample contains only about 63.2 % unique instances from the training set [BK99].
Boosting
Boosting was introduce by Schapire early 90’s as a method for boosting the performance
of a weak learning algorithm. The Adaboost algorithm (Adaptive Boosting), introduced
by Freund & Schapire [FS99], solve many of the practical difficulties of the earlier
boosting algorithms. The AdaBoost algorithm is given in Table 2.2. the algorithm takes
as input a training set ( ) ( )1 1, ,..., ,m mx y x y where each ix belongs to some domain or
CHAPTER 2 BACKGROUND
41
instance space X , and each label iy is in some label set Y . AdaBoost calls a given weak
or base learning algorithm repeatedly in a series of rounds1,...,t T= . One of the main
ideas of the algorithm is to maintain a distribution or set of weights over the training set.
The weight of this distribution on training sample i on roundt is denoted ( )tD i .
The weak learner’s job is to find a weak hypothesis { }: 1, 1ht X → − + appropriate for the
distribution Dt. The goodness of a weak hypothesis is measured by its error
( ): ( )
Pr ( )t
t i i
t i D t i i ti h x y
e h x y D i≠
= ≠ = ∑∼ (2.21)
In practice, the weak learner may be an algorithm that can use the weights tD on the
training samples. Alternatively, when this is not possible, a subset of the training
examples can be sampled according to tD , and these (unweighted) resampled examples
can be used to train the weak learner.
Table 2.2 the Boosting algorithm AdBoost. Adapted from [FS99]
( ) ( ) { }( )
{ }
1
Given: , ,..., , where , -1, 1
1Initialize
For 1, . . . , :
Train weak learner using distrubution
Get weak hypothesis : -1, 1 with error
i i
t
t
xi yi xm ym x X y Y
D i mt T
D
h X
∈ ∈ = +
=
=
•• → +
( )
1
Pr
11Choose ln .
2
Update:
if ( ) ( ) ( )
if ( )
t
t
t
t i D t i i
t
t
t i itt
t t i i
e h x y
e
e
e h x yD iD i
Z e h x y
α
α
α
∼
−
+
= ≠
−=
== × ≠
( ) ( )( )
1
exp =
where is a normalization factor (chosen so that will be a distribution).
Output the final hypothesis:
t t i t i
t
t t
D i y h x
Z
Z D
α
+
−
( )1
( ) .T
t tt
H x sign h xα=
= ∑
42
ExtraTrees
Geurts [Geu03] presents a new learning algorithm based on decision tree ensembles,
where the trees of the ensemble are built by selecting the tests during their induction fully
at random. This extreme randomization makes the construction of the ensemble very fast
even on very large datasets with high dimensionality.
The extra-trees algorithm builds an ensemble of unpruned decision or regression trees
according to the classical top-down procedure. Its two main differences with other tree-
based ensemble methods are that it splits nodes by choosing cut-points fully at random
and that it uses the whole learning sample (rather than a bootstrap replica) to grow the
trees.
Table 2.3 Extra-Trees splitting algorithm for numerical attributes [GEW06].
[ ]
( )
: the local learning subset corresponding to the node we want to split
: a split or nothing
- If ( ) is TRUE then return nothing.
- Otherwise select attribute
S
Input S
Output a ac
S
K
<
Split_a_node
Stop_split
{ }{ }
1
1
* * 1,
s , . . ., among all non constant (in ) candidate attributes;
- Draw splits , . . ., , where ( , ), 1, . . ., ;
- Return a split s such that Score( , ) max
K
K i i
i
a a S
K s s s S a i K
s S =
= ∀ ==Pick_a_random_split
...,
max min
Score( , ).
( , )
: a subset and an attribute
: a split
- Let and denote the maximal and minimal value of in ;
- Draw a random cut-point uniforml
K i
S S
c
s S
S a
Inputs S a
Output
a a a S
a
Pick_a_random_split
[ ]min min
min
y in , ;
- Return the split .
( )
: a subset
: a boolean
- If , then return TRUE;
- If all attributes are constant in , then return TRUE;
- If the output is co
S S
c
a a
a a
S
Input S
Output
S n
S
<
<
Stop_split
nstant in , then return TRUE;
- Otherwise, return FALSE
S
Table 2.3 shows the Extra-Tree procedure for numerical attributes. It has two parameters:
K, the number of attributes randomly selected at each node and nmin, the minimum
CHAPTER 2 BACKGROUND
43
smple size for splitting a node. It is used a several times with the (full) original learning
sample to generate an ensemble model (denoted by M the number of trees of this
ensemble). The predictions of the trees are aggregated to yield the final prediction, by
majority vote in classification problems and arithmetic average in regression problems.
Fuzzy Decision Trees
A fuzzy decision tree is a method able to partition the input space into a set of rectangles
and then approximate the output in each rectangle by a smooth curve, instead of a
constant or a class like in the case of crisp tree-based methods. A fuzzy tree is an
approximation structure to compute the degree of membership of objects to a particular
class (or concept) or to compute a numerical output of objects, as a function of the
attribute values of these objects. The goal is recursively split the input space into
(overlapping) subregions of objects which have the same membership degree to the target
class ( in the case of classification problems) or the same output value (in the case of
regression problems) [Ola04].
Figure 2.8 shows an example of splitting a fuzzy decision tree and its correspondent
graph.
1a2α 1α
1.0
0.0
1a
1a
11 2
βα< − 11 2
βα> +
22 2
βα< − 22 2
βα> +
0.8
0.50.2
1β2βˆCµ
Figure 2.8 Example of a fuzzy decision tree. Adapted from [Ola04].
A fuzzy decision tree structure is determined by the graph of the tree and by the attributes
attached to the its test nodes. The discretization thresholds (α ) and width (β ) values of
all these attributes, shown in Fig. 2.8, together the labels of all the terminal nodes
44
represent the parameters of the tree-based model. There is a search over both structure
and parameter spaces so as to learn a model from experience [Ola04].
2.2.3.5 Nearest Neighbor algorithm
The Nearest-Neighbor (1NN) method has been applied both for classification and
regression. Let an arbitrary instance x be described by the attribute vector
( ) ( ) ( )1 2, ,... na x a x a x , (2.22)
where ( )ra x denotes the value of the rth attribute of instance x . Then the distance
between two instances ix and jx is define to be ( ),i jd x x , where
( ) ( ) ( )( )2
1
,n
i j r i r jr
d x x a x a x=
≡ −∑ . (2.23)
In Nearest-Neighbor learning the target function may be either discrete-valued or real-
valued. Considering learning discrete-valued target functions of the form : nf Vℜ → ,
where V is the finite set { }1,..., sv v . The k-Nearest-Neighbor algorithm is shown in Table
2.4, the value ( )ˆqf x returned by this algorithm as its estimate of ( )qf x is just the most
common value of f among the k training examples nearest to qx . If we chose k=1, then
the 1-Nearest-Neigbor algorithm assigns to ( )ˆqf x the value of ( )qf x where xi is the
training instance nearest to qx . For larger values of k , the algorithm assigns the most
common value among the k nearest training sample [Mit97].
CHAPTER 2 BACKGROUND
45
Table 2.4 k-Nearest-Neighbor algorithm for approximation of a discrete value [MIT97].
( )Training algorithm:
For each training sample , , add the example to the list _
Classification algorithm:
Given a query instance to be classified,
Let q
x f x training examples
x
•
•
•
( ) ( )( )
1
1
. . . denote the instances from _ that are nearest to
Return
ˆ arg max ,
k q
k
qv V i
x x k training examples x
f x v f xiδ∈ =
•
← ∑
( ) ( ) where , 1 if and where , 0 otherwise.a b a b a bδ δ= = =
The main advantages of this algorithm are that it can in principle represent very complex
input-output relations very well and is very simple to implement. On the other hand, the
disadvantages can be numbered as follows:
• It does not handle many irrelevant attributes well. If we have lots of irrelevant
attributes, the distance between examples is dominated by the differences in these
irrelevant attributes and so becomes meaningless.
• It doesn’t look much like humans learning.
• Hypothesis function is too complex to describe explicitly.
• Computational inefficiency.
The nearest neighbor algorithm and its variants are particularly well suited to
collaborative filtering, where a system is to predict a given person’s preference based on
others people’s preferences. Collaborative filtering fits into the nearest neighbor search
well because attributes tend to be numeric and similar in nature, so it makes sense to give
them equal weight in the distance computation [Geu02].
46
2.2.4 Artificial neural networks (ANN)
2.2.4.1 Biological neural networks A neuron is a special biological cell that process information. It is composed of a cell
body and two types of out-reaching tree-like branches: the axon and the dendrites as
shown in Figure 2.9. A neuron receive signals (impulses) from others neurons through its
dendrites (Receivers) and transmits signals generated by its cell body along the axon
(transmitter), which eventually branches into strands and sub strands. At the terminal of
these strands are the synapses. A synapse is an elementary structure and functional unit
between two neurons (an axon strand of one neuron and a dendrite of another). When the
impulse reaches the synapse’s terminal, certain chemicals called neurotransmitters are
released. The neurotransmitters diffuse across the synaptic gap, to enhance or inhibit,
depending on the type of synapse, the receptor neuron’s own tendency to emit electrical
impulses. The synapse’s effectiveness can be adjusted by the signal passing through it so
that the synapses can learn from the activities in which their participate. This dependence
on history acts as a memory, which is possibly responsible for human memory [JMM96].
The cerebral cortex contains about 1011 neurons, this neurons are massively connected.
Each neuron is connected to 103 to 104 other neurons. In total, the human brain contains
approximately 1014 to 1015 interconnections.
Figure 2.9 A sketch of a biological neuron. Adapted from [JMM96.]
CHAPTER 2 BACKGROUND
47
2.2.4.2 Background
Artificial Neural networks (ANNs) are inspired by biological nervous systems and they
were first introduced as early as 1960. Nowadays, studies of ANNs are growing rapidly
for many reasons:
• ANNs work with pattern recognition at large
• ANNs have a high degree of robustness and ability to learn
• ANNs are prepared to work with incomplete and unforeseen input data
The development of artificial neural networks started several decades ago with the work
on the perceptron [Hay94]. The perceptron is basically a simple linear threshold unit, thus
able to represent only linear boundaries in the attribute space; its limited representation
capabilities have motivated the consideration of more complexes ANNs composed of
multiple interconnected layers of perceptrons [Weh98] and called Multi-Layer
Perceptrons (MLPs).
ANNs can be viewed as weighted directed graphs in which artificial neurons are nodes
are directed edges (with weights) are connections between neurons outputs and neurons
inputs. Based on the connection pattern (architecture), ANN’s can be grouped into two
categories [JMM96]:
• feed-forward networks, in which graphs have no loops.
• recurrent (or feedback) networks, in which loops occur because of feedback
connections.
An MLP is characterized by its architecture, training or learning algorithms and
activation functions. The architecture describes the connections between the neurons. It
consists of an input layer, an output layer and generally, one or more hidden layers in-
between. To each connection feeding hidden or output layers corresponds a weight.
These weights can then be adjusted to tune the input-output relationship of an MLP to
solve a given problem.
MLPs are normally used for supervised learning. In this context, the learning algorithm
makes use of both input-output data. Base on a set of input-output data, the weights are
48
updated so as to minimize the discrepancy between the given outputs and those computed
by the MLP from the given inputs. In this research, the most common algorithm,
backpropagation, is used. Backpropagation denotes actually an efficient algorithm for
computing the derivatives of the MLP output with respect to the weight values. It is used
as a main building block construct gradient descent or quasi-Newton algorithms to
minimize the discrepancy between MPL outputs and the desired ones provide in a
training sample..
Once trained, a network response can be, to a degree, insensitive to minor variations in its
input. This ability to see through noise and distortion to the pattern that lies within the
inputs is vital to pattern recognition in a real world environment.
A multi-layer network with one hidden layer is shown in Figure2.1010.
Input
Hidden layer
Output layer
Figure2.10 MLP with a single hidden layer
2.2.4.3The backpropagation method [Hay94,Nat97]
Backpropagation was created generalizing the Widrow-Hoff learning rule to multiple-
layer networks and nonlinear differentiable activation functions.
Standard back propagation is a gradient descent algorithm, as is the Widrow-Hoff
learning rule, in which the network weights are moved along the negative of the gradient
of the performance function. The term backpropagation refers to the manner in which the
gradient is computed or nonlinear multiplayer networks. There are a number of variations
on the basic algorithm that are based on other standard optimization techniques, such as
CHAPTER 2 BACKGROUND
49
conjugate gradient and Newton methods. Properly trained backpropagation networks tend
to give reasonable answers when presented with inputs that they have never seen.
Typically, a new input leads to an output similar to the correct output for inputs vectors
used in training that are similar to the new input being presented. The simplest
implementation of backpropagation learning updates the network weights and biases in
the direction in which the performance function decreases most rapidly – the negative of
the gradient
There are two different ways in which this gradient descent algorithm can be
implemented: incremental mode and batch mode. In the incremental mode, the gradient is
computed and the weights are updated after each input is applied to the network. In the
batch mode the weights and biases of the network are updated only after the entire
training set has been applied to the network. The gradients calculated at each training
example are added together to determine the change in the weights and biases. The batch
mode was used in this work.
The learning phase of a layered perceptron is where all its arc weight are adjusted
according to a specified learning rule in order to minimize a specified objective function
(energy function). A commonly used objective function E is the mean square error (MSE)
between the actual neural network outputs and the specified targets for a set of N training
patterns. The weight updating problem is to find a set of weights that minimizes the
predefined objective function [ElS96].
2.2.4.4 The Levenberg-Marquardt algorithm [HM94]
The Levenberg-Marquardt algorithm was designed to approach second-order training
speed without having to compute the Hessian matrix (the square matrix of second partial
derivatives of a scalar-valued function). When the performance function has the form of a
sum of squares (as is typical in training feed-forward networks), then the Hessian matrix
can be approximate as
JJH T= 2.22
and the gradient can be computed as
50
eJg T= 2.23
Where J is the Jacobian matrix that contains first derivatives of the network errors with
respect to the weights and biases, and e is a vector of network errors. The Jacobian
matrix can be computed trough a standard backpropagation technique that is much less
complex than computing the Hessian matrix.
Suppose that we have a function ( )V x which we want to minimize with respect to the
parameter vector x , then Newtons’s methods would be
( ) ( )12x V x V x−
∆ = − ∇ ∇ (2.24)
where ( )2V x∇ is the Hessian matrix, defined as follows [Zur92]:
( ) ( )2x x xV x V x ∇ = ∇ ∇ (2.25)
( )
2 2 2
21 1 2 1
2 2 2
22 1 2 22
2 2 2
21 2
n
nx
n n n
V V V
x x x x x
V V V
x x x x xV x
V V V
x x x x x
∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
∂ ∂ ∂ ∂ ∂ ∇ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
⋯
⋯
≜⋮ ⋮ ⋮
⋯
⋯
(2.26)
Note that the Hessian matrix is of size nxn and is symmetric. The matrix is often denoted
by H, thus ( )2V x∇ =H.
and ( )V x∇ is the gradient and is equal:
( )
1
2
n
V
x
V
xV x
V
x
∂ ∂
∂ ∂∇ ∂ ∂
≜
⋮
(2.27)
If we assumed that ( )V x is a sum of squared function
CHAPTER 2 BACKGROUND
51
( ) ( )2
1
N
ii
V x e x=
=∑ 2.28)
Then it can be shown that
( ) ( ) ( ) ( )2 TV x J x J x S x∇ = + (2.29)
( ) ( ) ( )TV x J x e x∇ = (2.30)
where ( )J x is the Jacobian matrix
( )
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
1 1 1
1 2
2 2 2
1 2
1 2
n
n
N N N
n
e x e x e x
x x x
e x e x e x
J x x x x
e x e x e x
x x x
∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
= ∂ ∂ ∂ ∂ ∂ ∂
∂ ∂ ∂
⋯
⋯
⋮ ⋮ ⋱ ⋮
⋯
(2.31)
and
( ) ( ) ( )2
1
N
i ii
S x e x e x=
= ∇∑ (2.32)
For the Gauss-Newton method it is assumed that ( ) 0S x ≈ , and update (2.24) becomes
( ) ( ) ( ) ( )1T Tx J x J x J x e x−
∆ = (2.33)
The Levenberg-Marquardt modification to the Gauss-Newton method is
( ) ( ) ( ) ( )1T Tx J x J x I J x e xµ−
∆ = + (2.34)
The parameterµ is multiplied by some factor ( )β whenever a step would result in an
increased ( )V x . When a step reduces ( )V x , µ is divided by β . Notice that when µ is
large the algorithm becomes steepest descent (with step 1 µ ), while for small µ the
algorithm becomes Gauss-Newton. The Levenberg-Marquardt algorithm can be
considered a trust-region modification to Gauss-Newton. Table 2.5 illustrates the
Levenberg-Marquardt modification to the backpropagation algorithm.
52
Table 2.5 Levenberg-Marquardt algorithm. Adapted from [HM94].
( )1) Present all inputs to the network and compute the corresponding network outputs,
and errors - . Compute the sum of squares of errors over all inputs (V(x)).
where is the error for the
Mq q q
q
e t a
e
=
qth input and is the output of the network when the qth input is presented.
2) Compute the Jacobian matrix.
3) Solve (2.34) to obtain x.
4) Recompute the sum of squares of errors using .
I
Mqa
x x
∆+ ∆
f this new sum of squares is smaller tahn taht computed in step 1, then reduce by ,
let , and go back to step 1. If the sum of squares is not reduced, then increase by
and go back
x x x
µ βµ β= + ∆
to step 3.
5) the algorithm is assumed to have converged when the norm of the gradient (2.30) is less
than some predetermined value, or when the sum of squares has been reduced to some error goal.
2.2.4.5 Feed-forward neural networks
Figure 2.11 depicts an example feed-forward neural network. A neural network can have
any number of layers, units per layer, network inputs, and network outputs. This network
has four units in the first layer (layer A) and three units in the second layer (layer B),
which are called hidden layers. This network has one unit in the third layer (layer C),
which is called the output layer. Finally, this network has four network inputs and one
network output. Some texts consider the network inputs to be an additional layer, the
input layer, but since the network inputs do not implement any of the functionality of a
unit, the network inputs will not be considered a layer in this discussion.
Network
inputs
Layer A
Units
Layer B
Units
Layer C
Units
Network input to
Unit connections
Unit to Unit
Connections
Network
Output
Figure 2.11 MLP with two hidden layers
CHAPTER 2 BACKGROUND
53
If a unit is in the first layer, it has the same number of inputs as there are network inputs;
if a unit is in succeeding layers, it has the same number of inputs as the number of units
in the preceding layer. Each network-input-to-unit and unit-to-unit connection (the lines
in Figure 2.11) is modified by a weight. In addition, each unit has an extra input that is
assumed to have a constant value of one. The weight that modifies this extra input is
called the bias. All data propagate along the connections in the direction from the
network inputs to the network outputs, hence the term feed-forward
CHAPTER 3 PROBLEM FORMULATION
54
CHAPTER 3 PROBLEM FORMULATION
3.1 ROTOR ANGLE AND SPEED ESTIMATION AND PREDICTION PROBLEM. It has been shown in the literature [SP89, APSZ95, MCK95, ABDA+96, CKP98,
LSTW99, LT95, AR96, SLT99, LT00, KGL01, KK02, BG04] that it is possible to
detect impending loss of synchronism by monitoring rotor angles and speeds of the
main power plants of an interconnection and that it is possible to use this information
to determine in real-time and in a closed loop fashion control actions in the form of
generator tripping, dynamic breaking or fast-valving. However, for these emergency
control schemes to work properly, it is necessary to obtain highly accurate estimates
of rotor angles and speeds. Also, to minimize the risk of detecting and reacting when
it is already too late (loss of synchronism may happen in some circumstances within
less than a few hundred milliseconds after fault inception), it is necessary to obtain
these estimates as quickly as possible, and if possible ahead in time via appropriate
prediction schemes.
Synchronous time frame rotor angles and speeds cannot be obtained easily by direct
measurements. On the other hand, their estimation can, at least in principle, be carried
out from the three-phase voltage and current phasors at the machine’s low voltage bus
[RLLM +95]. However, given the better accuracy of EHV phasor measurements with
respect to medium voltage ones, we suppose that the PMUs will in practice be
installed on the EHV side of the step up transformer of the power plant. Also, since
for transient instability monitoring it is not necessary to estimate angles and speeds of
each individual generator of a given power plant, we propose to use phasor
measurements from the EHV side of the step-up transformer to estimate and predict
only the rotor angle and speed of the COI of the considered power plant.
A PMU is a power system device that provides measurements of real-time phasors of
bus voltage and line currents. Basically, it samples (same time sampling) input
voltage and current waveforms using a common synchronizing signal from the global
positional satellite, GPS [Pha93], and calculates a phasor (modulus and angle) for the
fundamental frequency via Discrete Fourier Transform applied on a moving data
window whose width can vary from fraction of a sine wave cycle to multiple of the
CHAPTER 3 PROBLEM FORMULATION
55
cycle [Pha93,IEEE95]. These quantities are typically provided for the nodal voltage
and line and transformer currents in all three phases at the bus were the PMU is
connected.
In other words, the problem to be tackled can be formulated as follows:
Given, a time-series of three-phase voltage phasors and of out-flowing (three-
phase) current phasors acquired at the EHV side of a power plant, sampled at a
certain rate (typically one or two cycles), and represented in a synchronous
reference-frame at nominal frequency, compute an estimate of rotor angle and
speed of the centre of inertia of the power plant in the same reference-frame, and
compute a prediction of these quantities at the next and subsequent time-steps.
Our work will aim at showing the feasibility of such a local dynamic state estimation scheme for aggregated rotor dynamics of a power plant and we will also evaluate at the same time how long ahead in time it would be possible to predict angles and speeds.
3.2 APPROACH PROPOSAL TO SOLVE THE PROBLEM. The relationship between EHV PMU measurements and the dynamic state of the COI
of the power plant is essentially non-linear and typically corrupted by measurement
noise and modelling uncertainties. Therefore, we propose to use automatic learning
techniques, more specifically supervised multilayer perceptron training in order to
provide a black-box state estimation algorithm able to cope with such difficulties.
Indeed, it is well known that neural networks, and more generally automatic learning,
can cope with uncertainties and non-linearities, at least provided that the
dimensionality of their input space remains moderate. Note that this is the case in our
analysis, since typically the number of input variables will be in the range of a few
tens (at most 100) while by simulation it is possible to generate automatically a very
large sample of training scenarios (typically a few thousand). These scenarios can thus
cover a representative sample of power system configurations, fault scenarios,
modelling assumptions and they can also take into account measurement noise.
Training a neural network on such very large and representative scenarios thus may
presumably lead to a robust and at the same time very efficient state estimation
algorithm.
CHAPTER 3 PROBLEM FORMULATION
56
The approach investigated in this thesis thus essentially consists in generating off-line,
and based on numerical simulations, a representative training set composed of system
trajectories comprising inputs (sequences of voltage and current phasor
measurements) and output sequences of rotor angles and speeds of the COI of the
studied power plant. Obviously, it is of paramount importance that the set of
simulation scenarios is representative of all power system configurations and fault
scenarios.
It is clear that the neural network model needs to be updated when major changes
occur in the power system around the studied power plant, such as the installation of a
new transmission or generation equipment. On the other hand, since the relationship
between the COI of the power plant and PMU measurements depends on the number
of generators in operation in the plant, one suggestion is to train different neural
network models for each combination of generators in operation, and to use in real-
time the one corresponding to the actual configuration.
In order to evaluate the feasibility of the proposed approach we will carry out
experiments with two different power system models: the first one is a One-Machine-
Infinite-Bus (OMIB) system and the second one is a reduced version of the Mexican
Interconnected System (MIS). In our simulations, we have take care to use rather
detailed models of the power system dynamics and PMU device and we have
considered unbalanced as well as unbalanced conditions (e.g. due to single phase
faults).
CHAPTER 4 OMIB TEST SYSTEM.
57
CHAPTER 4 EXPERIMENTS WITH THE OMIB SYSTEM.
4.1 THE OMIB SYSTEM MODEL [Kun94] The OMIB test system, shown in Fig. 4.1, represents a thermal generating plant consisting of
four 555 MVA, 24 kV, 60 HZ units, connected to the rest of the system through a double
circuit transmission line. The equivalent machine is modeled with two damper windings in the
q-axis. The network reactances shown in Fig. 4.1 are in per unit on 2220 MVA, 24 kV base.
Resistances are assumed to be negligible and the infinite bus was modeled as an ideal 3-phase
AC voltage source (zero source impedance). Governor is not modeled in this example.
Mechanical power Pm is considered constant. Although very simple this system is very
helpful in understanding transient stability basic effects and concepts [Kun94].
0.5 . .j p u
0.93 . .j p u
0.15 . .j p u
Figure 4.1 OMIB test system (single-phase diagram). Adopted from [Kun94].
The parameters of the synchronous machine in per unit on 2220 MVA, 60 Hz base are
provided in Table 4.1.
The excitation system used is an IEEE standard type AC1A, and the parameters of this device
are given in Table 4.2 and 4.3 while the block diagram is illustrated in Figure 4.2.
CHAPTER 4 OMIB TEST SYSTEM.
58
Table 4.1 Synchronous Machine Parameters OMIB test system
Vbase 24 [kV] Ibase 13.3512312 [kA]
ω 376.991118 [rad/s] H 3.5 [MW/MVA] D 0.0 [p.u.] Ra 0.003 [p.u.] Ta 0.278 [sec] Xp 0.15 [p.u.] Xd 1.81 [p.u.] Xd' 0.30 [p.u.] Tdo' 8.0 [sec] Xd'' 0.23 [p.u.] Tdo'' 0.030 [sec] Xq 1.76 [p.u.] Xq' 0.65 [p.u.] Tqo' 1.0 [sec] Xq'' 0.25 [p.u.] Tqo'' 0.070 [sec]
Table 4.2 IEEE Alternator type AC1A Forward Path Parameters
Lead Time Constant (TC) 0.0 sec Lag Time Constant (TB) 0.0 sec
Regulator Gain (KA) 200 p.u. Regulator Time Constant (TA) 0.015 sec
MaxReg. Internal Volatge (VAMAX) 7.0 p.u. MinReg. Internal Voltage (VAMIN) -6.4 p.u. Max Regulator Output (VRMAX) 6.03 p.u. Min Regulator Output (VRMIN) -5.43 p.u.
Table 4.3 IEEE Alternator type AC1A Exciter Parameters
Rate Feedback Gain (KF) 0.03 p.u. Rate Feedback Time Constant (TF) 1.0 sec
Exciter time Constant (TE) 0.80 sec Exct. Constants related to fiel (KE) 1.0 p.u.
Saturation at VE1 0.1 p.u. Exciter Voltage for SE1 4.18 p.u.
Saturation at VE2 0.03 p.u. Exciter Voltage for SE2 3.14 p.u.
CHAPTER 4 OMIB TEST SYSTEM.
59
Figure 4.2 Exciter IEEE standard type AC1A. Adopted from [Hyd00]
4.2 DEVELOPMENT OF THE NEURAL NETWORKS FOR ROTOR ANGLE AND SPEED ESTIMATION The purpose of the ANNs is to estimate the rotor angle and speed of a synchronous machine
using voltage and current measurements, which are obtained by the PMU. We have trained
two different neural networks: one to estimate the rotor angle (ANN1) and another to estimate
the rotor speed (ANN2).
4.2.1 Input selection
The inputs to the neural network ANN1 are the voltage, current, angle of voltage and angle of
current at the EHV bus, at time instantst , 1t − and 2t − ,(where the time step is 16.66 ms),
totaling 12 inputs. The output of the neural network model consists of one neuron
representing the rotor angle for a specific operating condition,
( )
−−−−−−−−
=)(),(),(),(),(
),(),(),(),(),(),(),(
2t1tt2t1t
t2ti1titi2tv1tvtvft
iiivv
v
θθθθθθ
δ (4.1)
where )(tv and )(ti are the positive sequence terminal voltage and current at the time t ,
)( 1tv − , )2( −tv , )( 1ti − and )( 2ti − are the voltage and current at the time 1t − and 2t − , vθ and
iθ are the voltage and current angles at the same time instants.
On the other hand, for ANN2 we use the same inputs as with ANN1, with three inputs added,
the rotor angle obtained from the output of ANN1 at time instants t , 1t − and 2t − . For this
reason the number of inputs for ANN2 is 15. The output of the ANN2 consists of one neuron
representing the rotor speed as illustrated in Fig. 4.3.
CHAPTER 4 OMIB TEST SYSTEM.
60
vivθiθ
δ
ω
Figure 4.3 Arrangement of the ANNs for angle and speed estimation
4.2.2 Selection of ANN The ANNs used are of the multi-layer feed-forward type, with one hidden layer. Fig. 4.4
represents the multi-layer feed-forward network used for the purpose.
δδδδ
Input Layer Hidden Layer Output Layer
v(t-2)
v(t-1)
v(t)
i(t-2)
i(t-1)
i(t)
θθθθv(t-2)
θθθθv(t-1)
θθθθv(t)
θθθθi(t-2)
θθθθi(t-1)
θθθθi(t)
Figure 4.4 Proposed layered feed-forward ANN model for rotor angle estimation
The number of units in the hidden layer is determined experimentally, from studying the
network behavior during the training process taking into consideration some factors like
convergence rate, error criteria, etc. In this regard, different configurations were tested and the
best suitable configuration was selected based on the accuracy level required. The number of
hidden units for the ANN1 is 40 and the number of hidden units for ANN2 is 35. Hyperbolic
CHAPTER 4 OMIB TEST SYSTEM.
61
tangent activation functions are used for these units, while a linear activation function is used
for the output neurons for borh of ANNs . The neural networks were trained off-line.
4.3 SIMULATION RESULTS The Neural Network Toolbox from the MATLAB [Nat97] software tool was used to
create, train and test the neural networks. The training algorithm used is the Levenberg-
Marquardt algorithm because it provides fast convergence.
The initial weights as well as the initial biases employed random values between 0-1. The
inputs and targets are normalized so that they have values between –1 and 1.
A power system may be subjected to different kinds of disturbances. It is impossible to use all
the responses of the teaching system under different disturbances as the training set. The
contingencies represented in this well-kwon test system are three-phase short circuit at
beginning of the transmission line L2 and at the end of the same line near to infinite bus.
All the three-phase faults were applied at 0.1 sec. The faults were released either by self-
clearance or by tripping the faulted line. This is common practice in stability studies. All the
disturbances were applied to different generation levels [1100, 850, 600, 500, and 300 MW].
The training data uses 180 patterns, each containing 80 input-output pairs (in average). Total
number of input-output pairs is equal to 14400. To test the neural networks 60 unseen patterns
are used. Generation of the data for training and testing is summarized in Table 4.4. For each
short-circuit and generation level, 3 out of 9 patterns are with fault duration randomly chosen
from interval [0.05,CCT-0.01] ms, 3 from interval [CCT-0.01,CCT+0.01] ms, and 3 from
interval [CCT+0.01,0.35] ms.
Table 4.4 Generation of training and testing data.
Training Testing
Self-clearing fault
Tripping the line Self-clearing fault Tripping the line Gen. Level (MW) Beg.
of L2 End of
L2 Beg. of L2
End of L2
Beg. of L2
End of L2
Beg. of L2
End of L2
1100 9 9 9 9 3 3 3 3
850 9 9 9 9 3 3 3 3
600 9 9 9 9 3 3 3 3
500 9 9 9 9 3 3 3 3
300 9 9 9 9 3 3 3 3
CHAPTER 4 OMIB TEST SYSTEM.
62
Testing patterns consist of one pattern from all three, above mentioned, intervals that are not
used in training. All real-time environments exhibit some level of noise from instrumentation.
The effects of noise on the response of the system are assessed by randomly perturbing the
inputs (additive noise uniformly distributed in the range [-0.02,0.02]) to the neural networks.
The noise is added to voltage and current magnitude, only. First the ANN1 is trained and
tested, according to the procedure described above, and then the same training and testing
patterns are used with the ANN2
To generate the ANNs training and validation data sets, the MATLAB/ SIMULINK
software tool [Hyd00] is used. Also, using this simulation tool the values of voltage and
current phasors to compute the rotor angle and speed using the generator classical model,
were obtained. The sampling interval in the simulations is taken equal to 20 ms (every cycle
of fundamental frequency, this is reasonable value in view of the fact that modern PMUs are
capable to provide the measurements every 1-5 cycles [Tay00]).
As a measure of performance, the root mean square error defined as:
∑ −=p
2pp ot
p
1RMSE )( (4.2)
is determined for each of two ANNs after 1000 iterations of the training rule. In (4.2), p
represents the number of input-output training pairs, pt is the target output for the thp −
training, po is the output of the ANN. The RMSEs for training and testing are given in Table
4.5. For the comparison, the RMSEs obtained using the classical generator model for all three
presented cases are given in Table 4.6 (in equation (4.2) target output is replaced by exact
angle and speed values and the output of the ANN with the values obtained using the classical
generator model).
CHAPTER 4 OMIB TEST SYSTEM.
63
Table 4.5 Root mean square error after 1000 iterations
ANN Training error Testing error
ANN1 0.0020 (rad.) 0.0092 (rad.)
ANN2 0.0004 (rad./s) 0.0024 (rad./s)
Table 4.6 Root Mean Square Error for the Classical Generator Model
Results obtained for three cases (stable, critically stable, and unstable) are presented and
compared against the computation of the variables based on the classical generator model. All
three presented cases correspond to the faults at the beginning of the line L2 released by
opening the faulted line. CCT is equal to 0.292 seconds for this particular fault. If the fault
duration is less than the CCT, the system response is stable. The evolution of rotor angles and
speeds (exact, estimated, and obtained based on classical generator model) are illustrated in
Fig. 4.5 and 4.6. As the exact values of the rotor angles and speeds are considered those
extracted directly from the simulation model.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.60
0.5
1
1.5
2
2.5
time (sec)
roto
r angle
δ (ra
d)
exact classical modelANN1
Figure 4.5 Rotor angle (stable case)
CHAPTER 4 OMIB TEST SYSTEM.
64
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6-8
-6
-4
-2
0
2
4
6
time (sec)
roto
r spe
ed ω
(rad
/s)
exact classical modelANN2
Figure 4.6 Rotor speed (stable case)
An unstable system response (fault duration greater than the CCT) is illustrated in Fig. 4.7 and
4.8. When the fault duration is equal to the CCT system becomes critically stable. Fig. 4.9 and
4.10 represent the variables evolution for this case.
CHAPTER 4 OMIB TEST SYSTEM.
65
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
1
2
3
4
5
6
7
time (sec)
roto
r an
gle
δ (rad
)
exact ANN1 classical model
Figure 4.7 Rotor angle (unstable case)
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5-6
-4
-2
0
2
4
6
8
10
12
14
time (sec)
roto
r sp
eed
ω (rad
/s)
exact classical modelANN2
Figure 4.8 Rotor speed (unstable case)
CHAPTER 4 OMIB TEST SYSTEM.
66
Observe from Fig. 4.5, 4.7 and 4.9 that much better tracking of the rotor angle was obtained
by its estimation using the proposed methodology than if we rely on the classical generator
model and simple algebraic relations (1.1,1.2) see chapter 1 of this thesis. Presence of the
noise in measured variables results in slightly harsh aspect of rotor angle calculated by (1.1).
Rather harsh aspect in rotor speed is observable in all presented system responses if
analytical formulas (1.1,1.2) derived from the classical generator model are used. The harsh
aspects in rotor angle and speed are much less observable in the estimation using the ANNs.
If the level of accuracy, in transient stability assessment and control, is high then
observed errors in the computation of the variables using (1.1,.1.2) can result in wrong
prediction and control actions determination. The results clearly indicate that the ANN-
based approach to estimate rotor angles and speeds from phasor measurements, has potential
to be useful in tracking transient behavior of a power system following a disturbance.
0 0.5 1 1.50.5
1
1.5
2
2.5
3
3.5
time (sec)
roto
r an
gle
δ (rad
)
exact classical modelANN1
Figure 4.9 Rotor angle (critically stable case)
CHAPTER 4 OMIB TEST SYSTEM.
67
0 0.5 1 1.5-8
-6
-4
-2
0
2
4
6
8
time (sec)
roto
r sp
eed
ω (rad
/s)
exact classical modelANN2
Figure 4.10 Rotor speed (critically stable case)
68
CHAPTER 5 SIMULATIONS AND TRAINING RESULTS ON THE MEXICAN INTERCONNECTED SYSTEM
5.1 MODELS In this section we introduce the models used for the Mexican Interconnected System (MIS) that we have used in order to generate our simulation results. All the simulations reported in this thesis were carried out using the PSCAD/EMTDC software [Man03, Man03a].
5.1.1 Mexican interconnected system (MIS) General structure of the MIS The bulk Mexican interconnected system comprises a huge 400/230 kV transmission
system stretching from the border with Central America to its interconnection with USA.
The MIS consists of six areas designated as north (N), north-eastern (NE), western (W),
central (C), south-eastern (SE), and the peninsular systems. A simplified diagram of
major system elements is shown in Fig. 5.1.
Figure 5.1 Mexican Interconnected System. Adopted from [RRC97]
The test system used in our simulations is a reduced version of the MIS shown on Fig.
5.1. It is formed by one power plant which has 5 hydro-generators, three transmission
CHAPTER 5 SIMULATIONS AND TRAINING RESULTS ON THE MEXICAN INTERCONNECTED SYSTEM
69
lines that connect this power plant with the rest of the system, one load in the 400 kV bus
bar of the power plant and the rest of the system is represented by two large synchronous
machines and two large equivalent loads (see Figs. 5.3 and 5.6). A detailed model is used
to represent these latter two synchronous machines similar to the salient pole rotor one
used in the power plant. Excitation systems and governors are also modeled for all
synchronous machines. A standard IEEE exciter type AC1A was used as AVR model for
these two large synchronous machines..
Parameters of a part of South-East Mexican Interconnected System in PSCAD.
The synchronous machine represents a hydraulic (salient pole) generator model with one
damper winding in the q-axis. The set of equations that represent this model are given in
Appendix A and correspond to equations (A.10-A.17). Table 5.1 shows the parameters
used for each one of the five machines of the hydro-plant.
Table 5.1 Parameters of synchronous machine for the MIS test system (Power Plant)
Synchronous machine parameter for the MIS
T’ do 5.2 sec T’’ do 0.029 sec T’’ qo 0.034 sec
H 4.3 D 1.0 Xd 0.75 p.u; Xq 0.43 p.u. X’ d 0.24 p.u. X’’ d 0.17 p.u. X’’ q 0.17 p.u. X l 0.11 p.u.
The exciter is based on an IEEE type SCRX solid-state exciter. The schematic diagram is
shown in Fig. 5.2 and its parameter in Table 5.2.
70
Figure 5.2 IEEE Excitation system type SCRX [MAN03]
Table 5.2 Parameters of the excitation system SRCX
TA 1.02 (sec) TB 15.0 (sec) K 220.0 (p.u.) TE 0.03 (sec) EMIN -3.0 (p.u). EMAX 4.2 (p.u.)
The hydro-turbine is modeled using the same device shown in A.4. The parameters used
in our simulations are given in Tables 5.3 and 5.4.
Table 5.3 Hydro-Turbine Rated Conditions
Head at rated conditions 1.0 (p.u.) Output power at rated conditions 1.0, 0.90, 0.80 and 0.70 Gate position at rated conditions 1.0 (p.u.)
Rated No-load Flow 0.5 (p.u). Initial Output Power 1.0 (p.u).
Initial Operating Head 1.0 (p.u.)
Table 5.4 Hydro-Turbine Non-Elastic Water column parameters
Water Starting Time (TW) 2.26 (sec.) Penstock Head Loss Coefficient (fp) 0.02 (p.u.)
Turbine Damping Constant (D) 0.5
The hydro-governor model used is shown in A.5 and its parameters for this specific case
are shown in Table 5.5.
CHAPTER 5 SIMULATIONS AND TRAINING RESULTS ON THE MEXICAN INTERCONNECTED SYSTEM
71
Table 5.5 Hydro-Governor Parameters
Dead Value Band 0.0 (p.u.) Permanent Droop (Rp) 0.25 (p.u.)
Maximun Gate Position (Gmax) 0.8945 (p.u.) Munimun Gate Position (Gmin) 0.50 (p.u.)
Max Gate Opening Rate (MXGTOR) 0.17 (p.u./s) Min Gate closing Rate -0.17 (p.u./s)
Pilot Valve Servomotor Tiem Constant (Tp) 0.05 (sec) Servo gain (Q) 5 (p.u.)
Main Servo Time Constant 0.2 (sec) Temporary Droop (Rt) 0.4 (p.u.)
Reset or dashpottime constant 5.0 (sec) The step-up transformers have a delta-star configuration. Their model is based on the
theory of mutual coupling. Table 5.6 gives the parameters of each transformer.
Table 5.6 Transformer parameters for the MIS test sytem
Although our results would need to be further validated, by using a richer set of testing
scenarios, and also by taking into account measurement errors and variations in settings
of the power plant control loops, we believe that these conclusions will remain valid.
From a methodological point of view, although good results were obtained by using
MLPs, it would also be interesting to assess the possibility of using other supervised
learning methods, and possibly other sets of input variables.
CHAPTER 6 CONCLUSIONS AND FUTURE WORK
110
CHAPTER 6 CONCLUSION AND FUTURE WORK
6.1 CONCLUSIONS
This thesis focused on the estimation and prediction of rotor angles and rotor speeds of
synchronous machines using PMU measurements as inputs to artificial neural networks,
in order to train them over the data sets and evaluate their performances over independent
test sets.
We have chosen multi-layer perceptrons over others supervised learning methods because
this kind of algorithm is known to provide a good generalization provided enough
training data is available in comparison to the dimensionality of the input space. Ideed,
MLPs are smooth universal approximators, and we are working in a context where the
input space is of relatively low dimensionality while the target input-output mapping is
rather smooth. Another practically very important characteristic that has motivated the
use of MLPs is that, in spite of a rather slow off-line training algorithm, they can be used
in a very efficient manner to estimate or predict the quantities of interest in real-time.
This very high computational efficiency is particularly important in the context of
transient stability monitoring where it is necessary to reduce as much as possible delays
due to complex data processing.
In order to familiarize with the problem, we started our investigations by using a simple
test case, that consisted of an OMIB system. In this context, we used rather small data
sets of pairs of input-output data using a simplified dynamic model of the simulated
power system, implemented in MATLAB-SIMULINK. This model used a single phase
model of the system, assuming balanced conditions, and considered only the
electromechanical dynamics modeled in transient stability studies. This first set of
investigations used also the MATLAB neural network module and did not consider the
prediction of angles and speeds ahead in time. It yielded however promising results,
specially in comparison with an analytical approach exploiting the classical model to
CHAPTER 6 CONCLUSIONS AND FUTURE WORK
111
compute rotor quantities from electrical phasors. The results of this preliminary study
were published in a conference paper [DAGW03].
In a second stage, we have applied the same approach while considering a much more
detailed and realistic test system. In particular, to take into account the behavior of a
multi-machine system, we used a simplified model of the Mexican interconnected
system. This system consists basically of a power plant with five synchronous generators
represented in detail, that are interconnected by three lines to the rest of system. The rest
of the system is represented by two large synchronous machines in order to represent
with more fidelity the behavior of a real world system. In the simulations carried out on
this system we have also used a detailed ‘electromagnetic-transient and three-phase’
dynamical model and a more exact representation of the PMU device (modeling the FFT
used to compute phasors from instantaneous measurements). The higher complexity of
this problem obliged us to use a larger database of simulation scenarios to train the neural
networks. Hence we used the professional PEPITe [Pep04] data mining software in most
of the experiments on this system. In this study we found out that it was preferable to use
rectangular coordinates to represent the rotor angle targeted by the MLP [DAGE+06].
Further investigations have been reported in this thesis showing also quite promising
results concerning the possibility of predicting rotor angles and speed ahead of time.
These latter results have not yet been published.
6.2 FUTURE WORK
There are many possible directions for future work.
The most direct continuation would be to enhance the validation of the proposed
approach by considering other test power systems (in particular focusing on hydro-
plants), a more accurate representation of the PMU device (including measurement noise,
distortions to measurement transformers, and time jitter due to the limited accuracy of
GPS signals) and more extensive simulations on a broader range of system conditions.
A second direction of future research would consist in applying other supervised learning
methods to our datasets. In particular, we believe that it is worth comparing more
systematically the compromise between precision and computational requirements of
CHAPTER 6 CONCLUSIONS AND FUTURE WORK
112
multilayer perceptrons with other methods recently proposed in the automatic learning
literature (ensemble methods, kernel-based methods, support-vector machines etc.).
Also, since the long-term goal of our research is to enable real-time emergency control
for limiting the risk of loss of synchronism, we believe that it would be of interest to
investigate more deeply how automatic learning could be used in order to determine
directly the appropriate control actions. In this context, it would be particularly
interesting to investigate the possibility of determining these control actions by using
only locally acquired PMU measurements.
Finally, the approach investigated in this thesis should certainly be considered as a
candidate for other power systems instability monitoring questions, such as voltage
instability and negatively damped oscillations.
Bibliography
113
BIBLIOGRAPHY [ABDA+96] E. Abu-Al-Feilat, M. Bettayeb, H. Al-Duwaish, M. Abido and A. Mantawy, A
neural network-based approach for on-line dynamic stability assessment using synchronizing and damping torque, Electric Power System Research, Vol. 39, pp. 103-110, 1995.
[AF93] P. M. Anderson and A. A. Fouad, Power System Control and Stability, IEEE Press, 1993.
[ALA02] O. Anaya-Lara and E. Acha, Modelling and Analysis of Custom Power Systems by PSCAD/EMTDC, IEEE Transaction on Power Delivery, Vol. 17, No. 1, pp. 266-272, 2002.
[APSZ95] D. W. Aukland, I. E. D. Pickup, R. Shuttleworth and C. Zhou, Artificial neural network-based method for transient response prediction, IEE Proceedings Gen. Trans. Distribution, Vol. 142, No. 3, pp. 323-329,1995.
[AR96] F. Aboytes and R. Ramirez, Transient Stability Assessment in Longitudinal Power Systems using Artificial Neural Networks, IEEE Transactions on Power Systems, Vol. 11, No. 4, pp. 2003-2010, 1996.
[AW01] J. Arrillaga and N. A. Watson, Computer Modelling of Electrical Power Systems, John Wiley & Sons, 2001.
[BBBB92] N. Balu, T. Bertarnd et al, On-Line Power System Security, Proceeding of the IEEE, Vol. 80, No. 2, pp.262-280, 1992.
[BFOS84] L. Breiman, J.H. Friedman, R.A. Olshen, and C.J. Stone, Classification and regression trees, Wadsworth Int. 1984
[BG04] A. G. Bahbah and A. G. Girgis, New Method for Generator’s Angle and Angular Velocities Prediction for Transient Stability Assessment of Multimachine Power Systems Using Recurrent Artificial Neural Networks, IEEE Transactions on Power Systems, Vol. 19, No. 2, pp.1015-1022, 2004
[BG99] A.G. Bahbah and A. A. Girgis, Input feature selection for real-time transient stability assessment for artificial neural networks (ANN) using ANN sensitivy analysis, IN Proceedings of the 21 st IEEE International Conference of Power Industry Computer applications, PICA ’99, pp. 295-300, 1999.
[BK99] E. Bauer and R. Kohavi, An empirical comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants, Machine Learning, Vol. 36, pp. 105-139, 1999.
[BPA99] Bonneville Power Administration, WAMS Final Report, BPA, 1999.
[Bre96] L. Breiman, Bagging Predictors, Machine Learning, Vol. 24, pp. 123-140, 1996.
Bibliography
114
[ChZ97] T. J. Cholewo and J. M. Zurada, Sequential Network Construction for Time Series Prediction, In Proceedings of the IEEE International Joint Conference on Neural Networks, Houston, USA, pp. 2034-2039, 1997.
[CKP98] E. H. Camm, A. Keyhani and S. Pilluta, Developed of a neural network model for rotor angle estimation, Engineering Intelligent Systems, Vol. 6, No. 1, pp. 13-18, 1998.
[CLG00] R. Caruana, S. Lawrence and L. Giles, Overfitting in Neural Nets: Backpropagation, Conjugate Gradient, and Early Stopping, Neural Information Processing Systems, Denver, Colorado, 2000.
[CY96] Z. Chi and H. Yan. ID3-derived fuzzy-logic and optimized defuzzication for handwritten numerical recognition. IEEE Transactions on Fuzzy Systems, 4(1):24-31, February 1996.
[DAGE+06] A. Del Angel, , P. Geurts, D. Ernst ,M. Glavic and L. Wehenkel, Estimation of rotor angles of synchronous machines using artificial neural networks and local PMU-based quantities. Neurocomputing, available on line 27 February 2007.
[DAGW03] A. Del Angel, M. Glavic, L. Wehenkel, Using Artificial Neural Networks to Estimate Rotor Angles and Speeds from Phasor Measurements, In Porceedings of Intelligent System Applications in Power. ISAP 2003, Lemnos, Greece, Paper ISAP03/017, 2003.
[DEAS+04] I. C. Decker, J. G. Ehrensperger, M. N. Agostini, A.S. De Silva, A.L. Bettiol, S.L. Zimath, Synchronized Phasor Measurement System: Development and Applications, In proceeding of IX Symposium of Specialist in Electrical Operational and Expansion Planning (SEPOPE), paper SP-086, Rio de Janeiro, Brazil, 2004.
[DW02] A. Diu, L. Wehenkel, EXaMINE – Experimentation of a Monitoring and Control System for Managing Vulnerabilities of the European Infrastructure for electric Power Systems, In Proceedings of IEEE/PES Summer Meeting, Chicago, USA, 2002.
[ELO00] ELORSK: Wide Area Measurements of Power System Dynamics – The North American WAMS Project and its applicability to the Nordic Countries, Lund, Sweden, 2000.
[EP00] D. Ernst and M. Pavella, Closed-Loop Transient Stability Emergency Control, IEEE Power engineering Society Winter Meeting, Vol. 1, pp. 58-62,2000.
[Esh94] M. A. El-Sharkawi, What role can neural network play in power system engineering, IEEE Power Engineering Review, Vol. 14, No. 2, pp. 14-16,
Bibliography
115
1994.
[ESN96] M.A. El-Sharkawi and D. Niebur, Dynamic Security Assessment of Power Systems using Back Error Propagation Artificial Neural Networks, IEEE Power Engineering Society, 1996
[FS99] Y. Freund and R.E. Schapire, A short introduction to Boosting, Journal of Japanese Society for Artificial Intelligence, Vol. 14, No. 5, pp. 771-780, 1999.
[GBD92] S. Geman, E. Bienenstock, and R. Dousant. Neural Networks and the bias/variance dilemma. Neural Computation, 4:1-58, 1992.
[GDKI01] A. M. Gole, P. Demchenko, D. Krell and G.D. Irwin, Integrating Electromagnetic Transient Simulation with other Design Tools, In International Conference on Power Systems Transients, (IPST’2001), Rio de Janeiro, Brazil, 2001.
[Geu02] P. Geurts, Contributions to Decision Tree Induction: Bias/Variance tradeoff and time series classification. PhD thesis, University of Liege, Dept. of Electrical Engineering & Computer Science, Belgium, May 2002.
[Geu03] P. Geurts, Extremely randomized trees, Technical report, University of Liege, Department of Electrical Engineering and Computer Science, 2003.
[GEW06] P. Geurts, D. Ernst and L. Wehenkel, Extremely randomized trees, Machine Learning, Vol. 63, No. 1, pp. 3- 42, 2006.
[Hay94] S. Haykin, Neural Networks, A comprehensive Foundation, IEEE press, New York, 1994
[HHN05] K.E. Holbert, G. T. Heydt and H. Ni, Use of satellite technologies for Power System Measurements, Command and Control, Proceedings of the IEEE, Vol. 93, No. 5, pp. 947-955, 2005
[HKP90] J. Hertz, A. Krogh and R. G. Palmer, Introduction to the Theory of Neural Computation, Adison-Wesley Publishing Company, 1990.
[HM94] M. T. Hagan and M.B. Menhaj, Training Feedforward Networks with the Marquardt algorithm, IEEE Transaction on Neural Networks, Vol. 5 no. 6, pp. 989-993, 1994.
[Hyd00] Hydro-Quebec TEQSIM International, Power System Blockset for use with Simulink, User’s Guide version 2.0, The Matlab Works Inc., 2000.
[IEEE/CIGRE04]
IEEE/CIGRE Joint Task Force on Stability Terms and Definition, Definition and classification of Power System Stability, IEEE Transaction on Power Systems, Vol. 19, No. 2, pp. 1387-1401, may 2004.
[IEEE04] Task Force on Probalistic Aspects of reliability Criteria IEEE, Probabilistic Security assessment for power systems operation, Power Engineering Society,
Bibliography
116
General meeting, Vol. 1, pp. 212-220, 2004.
[IEEE95] IEEE 1344, Standard for Synchrophasors for Power Systems, New York, 1995
[JMM96] A. K. Jain , J. Mao and K.M. Mohiuddin, Artificial Neural Networks: A Tutorial, Computer, Vol. 29, No. 3, pp.31-44, 1996.
[KG02] I. Kamwa, R. Grondin, PMU configuration for system dynamic performance measurement in large multi-area Power Systems, IEEE Transaction on Power Systems, vol. 17, pp. 385-394, 2002
[KGH01] I. Kamwa, R. Grondin, Y. Hebert, Wide Area Measurement Based Stabilizing Control of Large Power Systems – A decentralized/Hierarchical Approach, IEEE Transactions on Power Systems, vol. 16, pp. 136-153, 2001 .
[KGL01] I. Kamwa, R. Grondin, L. Loud, Time-varing Contingency Screening for Dynamic Security Assessment Using Intelligent-Systems Techniques, IEEE Transactions on Power Systems, Vol. 16, No. 3, pp. 526-536, 2001.
[KH00] E.S. Karapidakis nad N. D. Hatziargyriou, Application of Artificial Neural Networks for Security Assessment of Medium Size Power Systems, 10th Mediterranean Electrotechnical Conference, MELECON 2000, vol. 3, pp. 1189-1192, 2000.
[Kim64] E. W. Kimbark, Power System Stability, Volume I Elements of Stability Calculations, John Wiley and Sons, 1964.
[KK02] G.G. Karady and M. A. Mohamed, Improving transient stability using generator tripping based on tracking rotor-angle and active power, Power Engineering Society Summer Meeting, Vol. 3, pp. 1576-1581, 2002.
[Kun04] P. Kundur, (convener). "Definition and Classification of Power System Stability." IEEE Transactions on Power Systems Vol. 19 ,No. 2: 1387-1401, 2004.
[Kun94] P. Kundur. Power System Stability and Control. McGraw Hill, 1994.
[KWS94] P.C. Krause, O.Wasynczuk and S. D. Sudhoff. Analysis of Electric Machinery.IEEE press, 1994.
[LSTW99] C. H. Liu, M. Ch. Su, S-S. Tsay and Y-J. Wang, Application of a novel fuzzy Neural Network to real-time transient stability swings prediction based on synchronized phasor measurement, IEEE Trans. on Power Systems, vol. 14, No. 2, pp. 685-692, 1999.
[LT00] C. W. Liu ans J.S. Thorp, New Methods for Computing Power System Dynamic Response for Real-Time Transient Stability Prediction, IEEE Transactions on Circuits and Systems-I: Fundamental Theory and Applications, Vol. 47, No. 3, pp. 324-337, 2000.
Bibliography
117
[LT95] C-W. Liu and J. Thorp, Application of synchronized Phasor measurements to real-time transient stability prediction, IEE Proceedings Generation, Trans. and Distribution, Vol. 142, No. 4, pp. 355-360, 1995.
[LTWS99] Ch. Liu, S. Tsay, Y. Wang and M. Su, Neuro-fuzzy approach to real-time transient stability prediction based on synchronized Phasor measurements, Electric Power Systems Research, Vol. 49, pp. 123-127, 1999.
[Man03] Manitoba HVDC Research Centre, PSCAD/EMTDC Power Systems Computed Aided Design, User’s guide on the use of PSCAD Simulation Software Tutorial, 2003
[Man03a] Manitoba HVDC Research Centre, EMTDC Transient Analysis for PSCAD Power System Simulation, User’s guide, A comprehensive resource for EMTDC, 2003
[MCK95] J. D. McCalley and B. A. Krause, Rapid transmission capacity margin determination for dynamic security assessment using artificial neural netwoks, Electric Power System Research,, Vol. 34, pp. 37-45, 1995.
[MCZ95] A. Malinowski, T. J. Cholewo and J. M Zurada, Capabilities and Limitations of Feedforward Neural Networks with Multilevel Neurons, In Proceedings of the IEEE International Symposium on Circuits and Systems, vo. 1, pp. 131-134, Seattle, USA, 1995.
[MEH00] J. A. Momoh, M. E. El-Havary, Electric Systems, Dynamics, and Stability with Artificial Intelligence Applications, 2000
[Mit97] T. M. Mitchell, Machine Learning, WCB/McGraw-Hill, 1997.
[Nat97] M. A. Natick, Neural Network Toolbox for use with Simulink user’s guide, The MathWorks Inc. 1997
[Ola04] A. C. Olaru, Contributions to Automatic Learning: Soft Decision Tree Induction, PhD thesis, University of Liege, Dept. of Electrical Engineering & Computer Science, Belgium, 2004.
[Pai81] M. A. Pai, Electric Power System Stability, Vol. 3 Analysis by the Direct Method of Lyapunov, North-Holland Publishing Company, 1981.
[Pep04] PEPITo User’s guide, 2004. www.pepite.be
[PERV00] M. Pavella, D. Ernst and D. Ruiz-Vega. Transient Stability of Power Systems: A UInified Approach to Assessment and Control. Kluwer Academic Publishers. 2000.
[Pha93] A.G. Phadke, Synchronized Phasor Measurements in Power systems, IEEE Computer Applications in Power, No. 6, pp. 10-15, 1993.
Bibliography
118
[PK96] Y. M. Park and W. Kim, Discrete-time adaptive sliding mode power system stabilizer with only input/output measurements, Electrical Power & Energy Systems, Vol. 18, No. 8, pp. 509-517, 1996.
[PM94] M. Pavella and P. G. Murthy, Transient Stability of Power Systems, John Wiley & Sons, 1994.
[PZ94] T. Parisini and R. Zoppoli, Neural Networks for Feedback Feedforward Nonlinear Control Systems, IEEE Transactions on Neural Networks, Vol. 5, No. 3, pp.436-449, 1994.
[RA05] M. A. Razi and K. Athappilly, A comparative predictive analysis of neural networks (NNs), nonlinear regression and classification and regression tree(CART) models, Expert Systems with Applications, vol. , No. , pp. 1-10, 2005.
[RLLM +95] S. Rovnyak, C-W. Liu, J. Lu, W. Ma and J. Thorp, Predicting future behavior of transient events Rapidly Enough to evaluate remedial control options in real time, IEEE Transaction on Power Systems, vol. 10, pp. 1195-1203, 1995.
[Sap90] G. SAPORTA, Probabilités analyse des données et statistique, Ed. Tecnip, 1990.
[SC02] K.S. Swarup, P.B. Corthis, ANN Approach Assess System Security, IEEE Computer Applications in Power, Vol. 15, No. 3, pp. 32-38,2002.
[SLT99] M. C. Su, C. W. Liu and S.S. Tsay, Neural-Network_based Fuzzy Model and its Application to Transient Stability Prediction in Power Systems, IEEE Transactions on Systems, Man, and Cybernetics-Part C: Applications and reviews, Vol. 29, No. 1, pp. 149-157, 1999.
[SP89] D.J. Sobajic and Y. H. Pao, Artificial Neural-Net Based Dynamic Security Assessment for Electrical Power Systems, IEEE Transactions on Power Systems, Vol. 4, No. 1, pp. 220-224,1989.
[SP98] P. W. Sauer and M. A. Pai, Power System Dynamics and Stability, Prentice-Hall, 1998.
[SVSH89] R. P. Schulz, L.S. VanSlyck and S. H. Horowitz, Applications of Fast Phasor Measurements on Utility Systems, Power Industry Application Conference PICA’89, pp. 49-55, 1989.
[Tay00] C. Taylor(Convener) CIGRE Task Force 38.02.17, Advanced Angle Stability Controls, CIGRE Technical brochure ,No. 155, 2000.
[TEMW+05] C. W. Taylor, D. C. Erikson, K. E. Martin, Robert E. Wilson and V. Venkatasubramanian, WACS-Wide-Area Stability and Voltage Control System: R&D and Online Demonstration, Proceedings of the IEEE, Vol. 93,
Bibliography
119
No. 5, pp.892-906, 2005.
[VH02] G. K. Venayagamoorthy and R. G. Harley, Two separate Continually Online-Trained Neurocontrollers for Excitation and Turbine Control of a Turbogenerator, IEEE Transactions on Industry Applications, Vol. 38, No. 3, pp.887-893, 2002.
[VHW02] G. K. Venayagamoorthy, R. G. Harley and D.C. Wunsch, Comparison of Heuristic Dynamic Programming and Dual Heuristic Programming Adaptive Critics for Neurocontrol of a Turbogenerator, IEEE Transactions on Neural Networks, Vol. 13, No. 3, pp. 764-773, 2002.
[VHW03] G. K. Venayagamoorthy, R. G. Harley and. D.C. Wunsch, Dual Heuristic Programming Excitation Neurocontrol for Generators in a Multimachine Power System, IEEE Transaction on Industry Applications, Vo. 39, No. 2, pp. 382-394, 2003.
[Weh96] L. Wehenkel, Contingency severity assessment for voltage security using non-parametric regression techniques, IEEE Transaction on Power Systems, Vol. 11, No. 1, pp. 101-111,1996.
[Weh98] L. Wehenkel, Automatic Learning techniques in power systems, Kulwer Academic Publishers, Boston, 1998. The effects of noise on the response of the system can be assessed by randomly perturbing the inputs (additive noise uniformly distributed).
[WLL05] Y. J. Wang, C. W. Liu and C. H. Liu, A PMU based special protection scheme: a case study of Taiwan power system, Electric Power Research, Vol. 27, No. , pp. 215-223, 2005.
[WREP05] L. Wehenkel, D. Ruiz-Vega, D. Ernst, M. Pavella, Preventive and Emergency Control of Power Systems, in Real-Time Stability in Power Systems – Techniques for Early Detection of the Risk of Blackout, S. Savulescu (ed.) Springer, pp. 199-232, 2005. [2]
[Zur92] J. M. Zurada, Introduction to Artificial Neural Systems, West Publishing Company, 1992.
Appendix A
120
Appendix A Modeling Power System Components
Generator model
A salient/round rotor Synchronous Machine fully developed model is available in
PSCAD/EMTDC. The model is programmed in state variable form, using generalized
machine theory.
The generalized machine model transforms the stator windings into equivalent commutador
windings, using the dq0 transformation as follows:
⋅
°−°−
°−°−
=
Vc
Vb
Va
Uo
Uq
Ud
2/12/12/1
)240cos()120sin()sin(
)240cos()120cos()cos(
θθθ
θθθ
(A.1)
The three-phase rotor winding may also be transformed into a two-phase equivalent winding,
with additional windings added to each axis to fully represent that particular machine, as is
shown in Figure A.1. Support subroutines are included in the machine model library for
calculating the equivalent circuit parameters of a synchronous machine from commonly
supplied data.
The d-axis equivalent circuit for the generalized machine is shown in figure A.2. Figure A.3
illustrates the flux paths associated with various d-axis inductances.
Figure A.1 Conceptual diagram of the three phase and dq windings
where,
Appendix A
121
k = Amortisseur windings
f = Field windings
a,b,c = Stator windings
d = Direct-axis ( d-axis) windings
q = Quadrature-axis (q-axis) windings
Figure A.2 D-axis Equivalent Circuit
LMD
L1
L23D
L3D L
2D
StatorAir Gap Amortisseur Field
Figure A.3 Flux Paths Associated with Various d-axis Inductances
A second damper winding on the q-axis is included and it can also be used as a round rotor
machine to model steam turbine generator.
Referring to Figures A.4 and A.5, the d-axis voltage UD2 and current ID2 are the field voltage
and current respectively. The damper circuit consist of parameters L3D and R3D with UD3=0.
The additional inductance L23D accounts for the mutual flux, which link only the damper and
field windings and not the stator windings. the following equations can be derived:
1 1 1 1
2 2 2 2
3 3 3 3
D q D D
D D D D D
D D D D
U R i id
U R i L idt
U R i i
ω ψ− ⋅ − ⋅
− ⋅ = ⋅ − ⋅
(A.2)
where,
Appendix A
122
+++
+++
+
=
DDMDDMDMD
DMDDDMDMD
MDMDMD
D
LLLLLL
LLLLLL
LLLL
L
32323
23223
1
(A.3)
( )32111 QQQMQQq iiiLiL ++⋅+⋅=ψ (A.4)
d
dt
θω =
(A.5)
Similar equations hold for the q-axis except the speed voltage term, dψυ ⋅ , is positive, and:
( )32111 DDDMDDd iiiLiL ++⋅+⋅=ψ (A.6)
Inversion of equation A.2 gives the standard state variable form BUAXX +=� with state vector
X consisting of the currents, and the input vector U, applied voltages. That is:
⋅+
⋅−
⋅−
⋅−⋅−
⋅=
−−
3
2
11
33
22
111
3
2
1
D
D
D
D
DD
DD
Dq
D
D
D
D
U
U
U
L
iR
iR
iR
L
i
i
i
dt
dψυ
(A.7)
⋅+
⋅−
⋅−
⋅−⋅−
⋅=
−−
3
2
11
33
22
111
3
2
1
Q
Q
Q
Q
QD
QD
Qd
Q
Q
Q
Q
U
U
U
L
iR
iR
iR
L
i
i
i
dt
dψυ
(A.8)
In the above form, Equation A.7 and A.8 are particularly easy to integrate. The equations are
solved using trapezoidal integration to obtain the currents. The torque equation is given as:
11 QdDq iiT ⋅−⋅= ψψ (A.9)
The dq-axis model includes the transient and sub transient characteristics of the machine and
the set of differential equations describing the generator dynamics is given by equations A.10
–A.15:
( ) ( )( )0 0
f q d d d q
qd d
E E Ef X X I EpE
T T
′ ′− + − −′ = =
′ ′ (A.10)
( )( )0 0
q q q dd
d
q q
X X I EEpE
T T
′ ′− − −−′ = =
′ ′ (A.11)
Appendix A
123
q q a q d dE V R I X I′′ ′′− = − (A.12)
d d a d q qE V R I X I′′ ′′− = − (A.13)
( )( )0
q d d d q
q
d
E X X I EpE
T
′ ′ ′′ ′′+ − −′′ =
′′ (A.14)
( )( )0
d q q q d
d
q
E X X I EpE
T
′ ′ ′′ ′′+ − −′′ =
′′ (A.15)
The final two state equations are provide by the rotor swing equations :
( )1
m e
dT T D
dt M
ωω= − − (A.16)
( 1)s
d
dt
δω ω= − (A.17)
where,
Tm – turbine torque
Te – electrical torque
M – inertia constant
D – Damping constant
ω - Machine speed
ωs – Synchronous speed
δ - Rotor angle
Governor
The governor can be represent by IEEE type thermal governor model. In this case the
approximate mechanical-hydraulic control (GOV1) was used. The schematic diagram for this
device is given in figure A.4
Appendix A
124
Figure A.4 hydro turbine model representation in PSCAD software.
Figure A.5 hydro governor model type IEEE GOV 1
Appendix B
125
Appendix B Phasor Measurements Units (PMU) Introduction. Phasors are basic tools of ac circuit analysis, usually introduced as a means of representing
steady state sinusoidal waveforms of fundamental power frequency.
Even when a power system is not quite in a steady state, phasor are often useful in describing
the behaviour of the power system. When the power system is undergoing electromechanical
oscillations during power swings, the waveforms of voltages and currents are not in steady
state, and neither is the frequency of the power system at its nominal value. Under these
conditions, as the variations of the voltages and currents are relatively slow, phasor may still
be used to describe the performance of the network, the variations being treated as a series of
steady state conditions. Recent developments in time synchronized techniques, coupled with
the computer based measurement technique, have provided a novel opportunity to measure
the phasors, and phase angle differences in real time.
Consider the steady state waveform of a nominal power frequency signal as shown in Figure
B.1. Starting to observe the waveform at the instant, the steady-state waveform may be
represented by a complex number with a magnitude equal to the RMS value of the signal and
with a phase angle equal to angleφ . In a digital measurement system, samples of the
waveform for one (nominal) period are collected according at 0t = , and the fundamental
frequency component of the Discrete Fourier Transform (DFT) is calculated according to the
relation:
2 /
1
2 Nj k N
kk
X XN
πε −
=
= ∑ , (B.1)
where N is the total number of samples in one period, X is the phasor, and kx is the wave
form samples. This definition of the phasor has the merit that it uses a number of samples
( )N of the wave form, and is the correct representation of the fundamental frequency
component, when other transient component are present [Pha93].
Phasors can be measured for each of the three phases (a, b, c), and the positive sequence
phasor can be computed according to its definition:
( )21
1
3 a b cX X X Xα α= + + (B.2)
Appendix B
126
where 2 /3j πα ε= .
φ
X
0t =
φ
X
Figure B.1 Phasor representation of a sinusoidal waveform. Adapted from [Pha03].
Synchronization signals could be distributed over any of the traditional communication media
currently in use in power systems. Most communication systems, such as leased lines,
microwave, or AM radio broadcasts, place a limit on the achievable accuracy of
synchronization, which is too coarse to be in practical use. Fibre-optic links could be used to
provide high precision synchronization signals, if a dedicated fibre is available for this
purpose. If a multiplexed fibre channel is used, synchronization errors of the order of 100
microseconds are possible, and are not acceptable for power system measurements. The
Geostationary Operational Environmental Satellites (GOES) systems have also been used for
synchronization purposes, but their performance is not sufficiently accurate [Wil92].
The technique of choice at present is the Navstar Global Positioning System (GPS) satellite
transmissions. This systems is designed primarily for navigation purposes, but if furnishes a
common-access timing pulse, which is accurate to within 1 microsecond at any location at
earth. The system uses transmissions from a constellation of satellites in non stationary orbits
at about 10 000 miles above the earth’s surface. For accurate acquisition of the timing pulse,
only one of the satellites need be visible to the antenna. The experience with the availability
and dependability of the GPS satellite transmission has been exceptionally good. [Pha93].
Appendix B
127
FFT Algorithm.
Fast Fourier Transform (FFT) is a classic filtering method. By performing Fourier transforms
over a window of N points, the frequency components /ek f N (where ef is the sampling
frequency) can be calculate using the formula below:
( )1
2
0
1ik
NN
j pi
i
X k v eN
−−
=
= ⋅∑ , (B.3)
where ie
iv v
f
=
.
More generally, ( )X k can be considered as the output from a filter which inputs ( )v t , with
exponential coefficients. We thus have:
( ) ( )1 2
10
1,
ikN j pN
n i Ni
X k X k n v eN
− −
+ − +=
= = ⋅ ⋅∑ (B.4)
The relationship between FFT and demodulation is thus expressed as follows:
( ) ( )11 2
10
11,
nN j pN
n i Ni
X k X n V eN
+− −
+ − +=
= = ⋅ ⋅
∑ (B.5)
The filter for cutting out the high frequency component consists in averaging over N points,
which cancels out frequencies which are multiples of ef N . If the sampling frequency is
right (i.e. if it is multiple of the frequency which contains the data signal) this filter will
remove the 02 f frequency together with the various harmonics [DCHH92].
Basic definitions.
The follow definition of a real-time or synchronized phasor is provided in the IEEE Standard
1344-1995 [IEEE95]:
• Anti-aliasing: By the Nyquist Theorem, the maximum reproducible frequency is one-
half the sampling rate. Aliasing is caused when frequencies higher than one-half of the
sampling rate are present. This results in the higher frequencies being ‘aliased’ down
to look like lower frequency components. Anti-aliasing is providing low pass filtering
to block out frequencies higher than those than ca be accurately reproduced by the
given sampling rate.
Appendix B
128
• Nyquist rate: the minimum rate that an analog signal must be sampled in order to be
represented in digital form. This rate is twice the frequency of that signal.
• Phase lock: the sate of synchronization between two ac signals in which they remain
at the same frequency and with constant phase difference. This term is typically
applied to a circuit that synchronized a variable oscillator with an independent signal
• Phasor: a complex equivalent of a simple sine wave quantity such that the complex
modulus in the sine wave amplitude and the complex angle (in polar form) is the sine
wave phase angle.
• Synchronism: the state where connected alternating-current systems, machines, or a
combination operate at the same frequency and where the phase angle displacement
between voltages in them are constant, or vary about a steady and stable average
value.
• Synchronized phasor: a phasor calculated from data examples using a standard time
signal as the reference for the sampling process. In this case, the phasors form remote
sites have a defined common phase relationship.
With real-time waveforms, it is necessary to define a time reference to measure phase angles
synchronously. The IEEE standard 1344-1995[IEEE95] defines the start of the second as the
time reference for establishing the phasor phase angle value.
The synchronized Phasor measurements convention is shown in Figure B.2
The instantaneous phase angle measurement remains constant at rated frequency when using
the start of the second phase reference. If the signal is at off-nominal frequency, the
instantaneous phase varies with time. The IEEE standard 1344-1995 defines a steady-state
waveform where the magnitude, frequency, and phase angle measurement performance for a
waveform do not change. This standard has no requirements regarding Phasor measurement
performance for a wave form in transient state [BSG04].
Appendix B
129
0φ = °
0V ∠ °
90V∠ − °90φ = − °
Figure B.2 Synchrophasor measurement convention with respect to time. Adopted from [BSG04].
Phasor Measuring Units.
The PMU was developed at the Virginia Polytechnic Institute, Blacksburg, in the mid-1980s.
The GPS time-synchronized PMU measures current and voltages in Phasor detail (i.e.
magnitude and phase). Phasor Measurements provide the capability to investigate power
system stability in greater detail [HHN05].
Phasor measuring units (PMU) using synchronization signals from the GPS satellite system
have evolved into mature tools and are now being manufactured commercially. Figure B.3
shows a typical synchronized Phasor measurement system configuration. The GPS
transmission is received by the receiver section, which delivers a phase-locked sampling clock
pulse to the analogue-to-digital converter system. The sampled data are converted to a
complex number which represents the Phasor of the sampled wave. Phasors of the three
phases are combined to produce the positive sequence measurement [BNKH+05].
Appendix B
130
The GPS receiver provides the 1 pulse-per-second (pps ) signal, and a time tag, which
consists of the year, day, hour, minute and second. The time could be the local time, or the
UTC (Universal Time Coordinate). The 1-pps signal is usually divide by a phase-locked
oscillator into the required number of pulses per second for sampling of the analog signals.
These signals are derived from the voltage and current transformer secondary sides, with
appropriate anti-aliasing and surge filtering.
Figure B.3 Phasor measurement unit. Adopte from [BNKH+05]
PMU representation in PSCAD [Man03a].
PSCAD include as a device the Fast Fourier Transform (FFT), Figure B.4 shows this device
in the PSCAD environment, which can determine the harmonic magnitude and phase of the
input signal as a function of time. The input signals first sampled before they are decomposed
into harmonic constituents. Options are provided to use one, two or three inputs. In the case
of three inputs, the component can provide output in the form of sequence components. In our
simulations we have selected the three 1-phase FFTs combined in one block. The input is
processed to provide the magnitudes Mag and phase angle Ph of the fundamental frequency
and its harmonics (including the DC component dc)
Figure B.4 FFT representation in PSCAD. Adopted from [Man03].
Appendix B
131
In Fig. B.4 the number 7 means the number of harmonics that FFT block calculates and its
imply that the number of samples per period of the fundamental frequency is set to be 16.
The task of frequency scanning involves a few data processing stages:
• Low-Pass Filtering (Anti-Aliasing)
• Sampling & Fourier Transform
• Phase and Magnitude Error Correction.
Figure B.5 illustrates graphically this process inside the FFT block.
Figure B.5 On-line frequency scanner in PSCAD/EMTDC.
Computations are performed on-line, at each sampling instance, and are based on a sampled
data window of the preceding input signal cycle. In accordance with the Nyquist Criteria,
data sampling is performed at a frequency greater than double the highest harmonic frequency
of interest. Sampling rates may be one of, 16, 32, 64, 127 or 255 samples/cycle of
fundamental frequency, which are written to a buffer. In our simulations 16 samples/cycles is
selected.
Since the number of samples in a window represents a period of fundamental frequency, the
dynamics of a cycle preceding a sample are captured in the computations. It should be noted
that outputs of this subroutine contain valid information only if a complete data window is
available for computations
It is important to be aware of the inherent aliasing effects due to sampling of the input signal.
A low pass, anti-aliasing filter is recommended at all times, unless the input signal is
guaranteed not to have any higher order harmonics. This filter is provided within the
component.
The harmonic computations are based on a standard Fast Fourier Transformation (FFT)
technique, used in digital signal processing. The basis function for computation of phase
Appendix B
132
angle can either be a fundamental frequency cosine waveform or a sine waveform starting at
time = 0.
The harmonics computed are with respect to a given constant fundamental frequency. For
situations where the fundamental frequency is variable, the use of a frequency-tracking device
is available to the user. The frequency-tracking unit uses the fundamental component of the
input signal corresponding to the previous sampling instance (as computed by the FFT
routine), to monitor small changes in the frequency of the input signal. This element is meant
to monitor minor fluctuations of frequency. Frequency tracking may be enabled or disabled at
users discretion.
Gibbs ringing effect, as a result of rectangular data windows, is usually not a problem with
harmonics of the fundamental frequency. However, if the sampling frequency is not
synchronized to the fundamental frequency of the input signal, the Gibbs effect distortions
introduced on the measurement of harmonics may be significant. Therefore, use of the
frequency-tracking feature may not be needed unless the fundamental component is